Learning to Verify: Signature-Guided Training for Adaptive Adversarial Detection
Summary
The paper proposes a signature-guided dual-model verifier that keeps the classifier fixed while training a separate verifier using activation signatures to detect adversarial examples under adaptive attacks, outperforming confidence- and feature-based detectors on L∞ attacks but showing limitations on L2 perturbations.
View Cached Full Text
Cached at: 11/20/26, 09:16 AM
# Learning to Verify: Signature-Guided Training for Adaptive Adversarial Detection
Source: [https://arxiv.org/html/2611.25271](https://arxiv.org/html/2611.25271)
Roozbeh Razavi\-FarAffiliation:Trustworthy and Secure AI \(TSAI\) Lab, Faculty of Computer Science,Affiliation:University of New Brunswick, Fredericton, NB, CanadaAffiliation:arash\.vashagh@unb\.ca, roozbeh\.razavi\-far@unb\.ca
###### Abstract
Adversarial example detectors often fail when attackers explicitly attempt to bypass them\. We propose a dual\-model approach that keeps the original classifier fixed and trains a separate verifier to assess its predictions\. During training, we identify patterns of internal neural network activations that distinguish genuine examples of each class from adversarial examples targeting that class\. These patterns, called activation signatures, guide the generation of challenging adversarial examples that mislead the classifier, challenge the verifier, and internally resemble the target class\. The verifier learns to recognize these examples while preserving its performance on clean inputs through knowledge distillation\. At inference, the verifier evaluates the classifier’s prediction without requiring activation signatures\. We evaluate the approach on publicly available datasets against several adaptive attacks, including attacks unseen during training\. The proposed verifier substantially outperforms directly attacked confidence\-based and feature\-based detectors under the evaluatedL∞L\_\{\\infty\}attacks\. Ablation studies show that adaptive verifier training is the main source of this improvement, while signature guidance primarily benefits clean accuracy\. AdditionalL2L\_\{2\}evaluations reveal limitations in robustness across perturbation types\.
###### Index Terms:
adversarial example detection, adversarial robustness, adaptive attacks, activation signatures, dual\-model verification, adversarial training, and adversarial machine learning\.
## IIntroduction
Deep neural networks achieve strong performance in image classification, but small input perturbations can cause confident but incorrect predictions\[[1](https://arxiv.org/html/2611.25271#bib.bib24)\]\. Adversarial training improves robustness by exposing a model to perturbed examples during optimization\[[2](https://arxiv.org/html/2611.25271#bib.bib23)\], but its protection may depend on the attacks used during training, and retraining may reduce clean accuracy\. Adversarial example detection provides an alternative by rejecting inputs that do not appear trustworthy\. The problem becomes harder under adaptive attacks, where the attacker jointly optimizes misclassification and detector evasion\[[3](https://arxiv.org/html/2611.25271#bib.bib16),[4](https://arxiv.org/html/2611.25271#bib.bib18)\]\.
Existing detectors use classifier confidence, energy, feature\-space distances, intermediate activations, or complementary prediction and rejection models\[[5](https://arxiv.org/html/2611.25271#bib.bib30),[6](https://arxiv.org/html/2611.25271#bib.bib31),[7](https://arxiv.org/html/2611.25271#bib.bib28),[8](https://arxiv.org/html/2611.25271#bib.bib26),[9](https://arxiv.org/html/2611.25271#bib.bib1)\]\. However, fixed scores and simple model agreement can remain vulnerable when the attacker knows the detection rule\. Detectors must also generalize beyond the attack procedures observed during training\.
We study whether a separate verifier can provide an adaptive rejection signal without modifying the deployed classifier\. The verifier should learn from attacks that challenge its decision rule while remaining simple at inference\. A reliable evaluation should attack each detector directly, include unseen attacks, verify perturbation constraints, and report attack success\.
We propose a signature\-guided dual\-model verifier in which a fixed classifier produces the prediction and a separately trained verifier measures support for that prediction\. Our contribution combines detector\-aware training of a separate verifier, a training\-only signature curriculum, and prediction\-conditioned rejection without a test\-time signature bank\. The signatures help construct examples that fool the classifier, challenge the verifier, and internally resemble the target class\. The verifier learns to recover the original label while preserving clean behavior through label smoothing and distillation\. During inference, the signature bank is removed, and detection uses the verifier’s probability for the classifier’s predicted class with a class\-specific threshold\. Deployment requires one forward pass through each model\.
We evaluate the method on CIFAR\-10, SVHN, and STL\-10 using three random seeds\. The evaluation includes detector\-aware PGD and APGD, a CW\-styleL2L\_\{2\}attack, and AutoAttack candidates followed by detector\-aware refinement\. We also directly attack MSP, energy, diagonal Mahalanobis distance, and an activation\-based logistic detector using their respective scores\. The full verifier outperforms these baselines and behaves similarly across the threeL∞L\_\{\\infty\}procedures, including APGD and AA\+Refine, which are unseen during the verifier training\. AA\+Refine generates adversarial candidates using AutoAttack\[[10](https://arxiv.org/html/2611.25271#bib.bib19)\], and then, applies detector\-aware PGD refinement\. We further compare a no\-signature verifier and evaluate DenseNet121 and directL2L\_\{2\}PGD\. These comparisons distinguish the value of adaptive verifier training from the measured effect of signature guidance on clean predictions\. The implementation is available at[https://github\.com/arashVsh/learning\-to\-verify\-ssci\-2027](https://github.com/arashVsh/learning-to-verify-ssci-2027)\.
## IIBackground
This section reviews robust learning, adversarial detection, rejection, and verification methods, as well as adaptive evaluation\. Recent surveys provide a broader view of adversarial machine learning by organizing attacks according to attacker knowledge, objectives, capabilities, and learning stage, and by grouping defenses into several strategies\[[11](https://arxiv.org/html/2611.25271#bib.bib32),[12](https://arxiv.org/html/2611.25271#bib.bib33)\]\. Within this broader landscape, our work focuses on test\-time evasion detection and on evaluation against adaptive white\-box attacks\.
### II\-AAdversarial Examples and Robust Learning
Adversarial training minimizes loss under perturbations generated within a predefined threat model\[[2](https://arxiv.org/html/2611.25271#bib.bib23)\]\. Its protection may weaken when the test\-time attack differs from the training attack, which limits conclusions drawn from a single perturbation model\. Continual robust training addresses evolving adversaries without repeatedly rebuilding the model\[[13](https://arxiv.org/html/2611.25271#bib.bib8)\]; more broadly, self\-adaptive continual learning addresses harsh, changing conditions\[[14](https://arxiv.org/html/2611.25271#bib.bib35)\]\. Decision\-based methods further demonstrate that adversarial examples can be produced without access to gradients or model parameters\[[15](https://arxiv.org/html/2611.25271#bib.bib14)\]\. This broad range of attack settings shows that robustness must be evaluated beyond one training algorithm or access assumption\.
### II\-BAdversarial Example Detection
Early detection methods train auxiliary networks using intermediate features, while later robust objectives incorporate adversarial examples into detector training\[[8](https://arxiv.org/html/2611.25271#bib.bib26),[16](https://arxiv.org/html/2611.25271#bib.bib27)\]\. Other approaches use class\-conditional feature distances, unusual activation subsets, neural fingerprints, or fuzzy rules\[[7](https://arxiv.org/html/2611.25271#bib.bib28),[17](https://arxiv.org/html/2611.25271#bib.bib29),[18](https://arxiv.org/html/2611.25271#bib.bib2),[19](https://arxiv.org/html/2611.25271#bib.bib25)\]\. These methods differ in how they represent abnormal behavior, but most depend on a score learned or calculated from the classifier’s internal response\. Layer\-wise responses and graph\-guided model behavior provide additional signals by capturing where perturbations affect the network and how model behaviors relate\[[20](https://arxiv.org/html/2611.25271#bib.bib6),[21](https://arxiv.org/html/2611.25271#bib.bib3)\]\. Robust anomaly detection and sequential statistical testing instead address weaknesses in learned negative examples and violations of exchangeability\[[22](https://arxiv.org/html/2611.25271#bib.bib5),[23](https://arxiv.org/html/2611.25271#bib.bib10)\]\. Recent work has also explored transferring adversarial detectors across different classifier architectures\. GraphRectify uses graph\-based representations of intermediate features to adapt detectors between neural network backbones\[[24](https://arxiv.org/html/2611.25271#bib.bib34)\]\.
Output\-based scores remain useful baselines because they require no additional feature extractors\. MSP identifies unusually weak classifier confidence, while energy uses the complete logit vector rather than only its largest probability\[[5](https://arxiv.org/html/2611.25271#bib.bib30),[6](https://arxiv.org/html/2611.25271#bib.bib31)\]\. Energy\-based modeling has also been connected to scalable adversarial training\[[25](https://arxiv.org/html/2611.25271#bib.bib15)\]\. These inexpensive scores may nevertheless miss adversarial inputs that receive confident incorrect predictions\.
### II\-CVerification, Rejection, and Generalization
Rejecting an input can be preferable when the evidence for a prediction is insufficient\. Two\-head models combine prediction and rejection, agreement checking compares model behavior, and reject\-option learning balances clean accuracy, robustness, and abstention\[[9](https://arxiv.org/html/2611.25271#bib.bib1),[26](https://arxiv.org/html/2611.25271#bib.bib7),[27](https://arxiv.org/html/2611.25271#bib.bib11)\]\. Selective\-classification theory further studies when abstention is preferable to a potentially incorrect decision\[[28](https://arxiv.org/html/2611.25271#bib.bib13)\]\. In this setting, a model is allowed to withhold a prediction on uncertain inputs, so performance is considered jointly in terms of prediction quality and the fraction of inputs on which the model chooses to decide\.
Detectors must also recognize attacks not observed during the training session\. Few\-shot approaches use limited attack examples to reduce reliance on large labeled collections\[[29](https://arxiv.org/html/2611.25271#bib.bib4)\]\. Distillation offers another way to transfer predictive structure, with robust variants encouraging a student to reproduce useful teacher behavior under perturbations\[[30](https://arxiv.org/html/2611.25271#bib.bib9),[31](https://arxiv.org/html/2611.25271#bib.bib12)\]\. These studies motivate verification systems that use complementary model behavior rather than one confidence estimate\.
### II\-DAdaptive Evaluation
A white\-box attacker has access to the detector’s decision rule and can optimize classification success and detector evasion jointly\[[3](https://arxiv.org/html/2611.25271#bib.bib16),[4](https://arxiv.org/html/2611.25271#bib.bib18)\]\. Detectors may appear effective against fixed classifier attacks yet fail once their own score is included in the objective\. Attacks that cannot optimize through a defense may also create misleading robustness through gradient masking\[[32](https://arxiv.org/html/2611.25271#bib.bib17)\]\. Diverse attack ensembles and diagnostic indicators help to distinguish genuine robustness from failed optimization\[[10](https://arxiv.org/html/2611.25271#bib.bib19),[33](https://arxiv.org/html/2611.25271#bib.bib20)\]\. Detection can remain as difficult as robust classification under a fully informed adversary, and test\-time defenses require evaluation of their complete adaptive behavior\[[34](https://arxiv.org/html/2611.25271#bib.bib21),[35](https://arxiv.org/html/2611.25271#bib.bib22)\]\. A credible evaluation should therefore attack each detector directly, verify perturbation limits, and report attack success so the detector’s performance is not attributed to weak attacks\.
## IIIMethodology
This section describes signature construction, signature\-guided hard\-example generation, verifier training, and the final prediction\-conditioned rejection rule\. Figure[1](https://arxiv.org/html/2611.25271#S3.F1)summarizes the proposed verifier\.
Fig\. 1:Overview of the proposed method\. Activation signatures guide hard\-example generation during the training session\. During inference, the fixed classifier and hardened verifier provide the prediction and verification score used by the class\-specific rejection rule\.Let𝒞=1,…,C\\mathcal\{C\}=\{1,\\ldots,C\}denote the set ofCCclasses\. The labeled training set is divided into four disjoint partitions\. The signature, verifier, development, and calibration partitions are denoted by𝒟sig\\mathcal\{D\}\_\{\\mathrm\{sig\}\},𝒟ver\\mathcal\{D\}\_\{\\mathrm\{ver\}\},𝒟dev\\mathcal\{D\}\_\{\\mathrm\{dev\}\}, and𝒟cal\\mathcal\{D\}\_\{\\mathrm\{cal\}\}, respectively\. They contain 20%, 50%, 15%, and 15% of the training samples with class\-level stratification\. The base classifier is trained using𝒟sig∪𝒟ver\\mathcal\{D\}\_\{\\mathrm\{sig\}\}\\cup\\mathcal\{D\}\_\{\\mathrm\{ver\}\}\. Signature construction uses𝒟sig\\mathcal\{D\}\_\{\\mathrm\{sig\}\}\. Verifier optimization uses𝒟ver\\mathcal\{D\}\_\{\\mathrm\{ver\}\}, while checkpoint selection and threshold calibration use𝒟dev\\mathcal\{D\}\_\{\\mathrm\{dev\}\}and𝒟cal\\mathcal\{D\}\_\{\\mathrm\{cal\}\}\. The official test set𝒟test\\mathcal\{D\}\_\{\\mathrm\{test\}\}is reserved for final evaluation\.
The fixed classifier is denoted byf𝜽f\_\{\\boldsymbol\{\\theta\}\}\. Its parameters are represented by𝜽\\boldsymbol\{\\theta\}\. The verifier is denoted bygϕg\_\{\\boldsymbol\{\\phi\}\}\. Its trainable parameters are represented byϕ\\boldsymbol\{\\phi\}\. Both models return vectors of logits\. The classifier probability distribution isπ𝜽\(⋅∣𝐱\)=softmax\(f𝜽\(𝐱\)\)\\pi\_\{\\boldsymbol\{\\theta\}\}\(\\cdot\\mid\\mathbf\{x\}\)=\\operatorname\{softmax\}\(f\_\{\\boldsymbol\{\\theta\}\}\(\\mathbf\{x\}\)\)\. The verifier probability distribution isπϕ\(⋅∣𝐱\)=softmax\(gϕ\(𝐱\)\)\\pi\_\{\\boldsymbol\{\\phi\}\}\(\\cdot\\mid\\mathbf\{x\}\)=\\operatorname\{softmax\}\(g\_\{\\boldsymbol\{\\phi\}\}\(\\mathbf\{x\}\)\)\. The input is denoted by𝐱\\mathbf\{x\}and its true label isy∈𝒞y\\in\\mathcal\{C\}\.
### III\-ASignature\-Guided Verifier Training
We construct one dataset\-level signature bank with a class\-conditioned entry for every class\. Let𝐚\(𝐱\)\\mathbf\{a\}\(\\mathbf\{x\}\)denote the activation vector formed by flattening and concatenating outputs from selected late layers off𝜽f\_\{\\boldsymbol\{\\theta\}\}\. The implementation extracts activations from the last eight layers whose parameter counts exceed a fixed significance threshold\.
Clean activations are collected from𝒟sig\\mathcal\{D\}\_\{\\mathrm\{sig\}\}\. TargetedL∞L\_\{\\infty\}PGD examples are also generated, but an example is retained only whenf𝜽f\_\{\\boldsymbol\{\\theta\}\}predicts its predetermined target\. Target labels are selected before optimization and balanced across classes\. If the finite guidance partition cannot provide the requested quota, the largest available balanced pool is retained as long as every target reaches at least half of that quota\.
For each classc∈𝒞c\\in\\mathcal\{C\}, mutual information identifies coordinates that distinguish genuine examples of classccfrom successful adversarial examples targetingcc\. A variance filter first removes coordinates with little variation\. The implementation retains at most 8192 variance candidates, and, then, selects at most 256 coordinates for each class\.
Letℐc\\mathcal\{I\}\_\{c\}denote the set ofKKcoordinates selected for classcc\. For coordinatej∈ℐcj\\in\\mathcal\{I\}\_\{c\}, the median clean activation isμcj\\mu\_\{cj\}\. Its robust scale isρcj\\rho\_\{cj\}, which is calculated as the median absolute deviation multiplied by 1\.4826\. A positive floor is applied toρcj\\rho\_\{cj\}to prevent unstable normalization\.
The smoothL1L\_\{1\}function is denoted byhh\. Its quadratic branch ish\(q\)=q2/2h\(q\)=q^\{2\}/2for\|q\|<1\|q\|<1\. Its linear branch ish\(q\)=\|q\|−1/2h\(q\)=\|q\|\-1/2for\|q\|≥1\|q\|\\geq 1\. The robust distance between𝐱\\mathbf\{x\}and the signature of classccis
dc\(𝐱\)=1K∑j∈ℐch\(aj\(𝐱\)−μcjρcj\),d\_\{c\}\(\\mathbf\{x\}\)=\\frac\{1\}\{K\}\\sum\_\{j\\in\\mathcal\{I\}\_\{c\}\}h\\left\(\\frac\{a\_\{j\}\(\\mathbf\{x\}\)\-\\mu\_\{cj\}\}\{\\rho\_\{cj\}\}\\right\),\(1\)whereaj\(𝐱\)a\_\{j\}\(\\mathbf\{x\}\)is coordinatejjof the activation vector\. A smaller distance indicates greater similarity to the clean activation pattern of classcc\.
An input may be close to several class signatures\. Therefore, the selected class is compared with every competing entry in the bank\. Letγ\>0\\gamma\>0denote the soft\-minimum temperature\. The competing\-class distance is
oc\(𝐱\)=−γlog∑k∈𝒞,k≠cexp\(−dk\(𝐱\)γ\)\.o\_\{c\}\(\\mathbf\{x\}\)=\-\\gamma\\log\\sum\_\{k\\in\\mathcal\{C\},\\,k\\neq c\}\\exp\\left\(\-\\frac\{d\_\{k\}\(\\mathbf\{x\}\)\}\{\\gamma\}\\right\)\.\(2\)
Intuitively, the soft minimum emphasizes the nearest competing class while retaining gradients from the others\. The relative signature margin ismc\(𝐱\)=dc\(𝐱\)−oc\(𝐱\)m\_\{c\}\(\\mathbf\{x\}\)=d\_\{c\}\(\\mathbf\{x\}\)\-o\_\{c\}\(\\mathbf\{x\}\)\.
A negative margin indicates that the input is closer to classccthan to its strongest competitor\. Distance and margin are standardized before they are combined\. The resulting cross\-bank geometry penalty is
G\(𝐱,c\)=h\(dc\(𝐱\)−d¯cscd\)\+h\(mc\(𝐱\)−m¯cscm\),\\displaystyle G\(\\mathbf\{x\},c\)=h\\left\(\\frac\{d\_\{c\}\(\\mathbf\{x\}\)\-\\bar\{d\}\_\{c\}\}\{s\_\{c\}^\{d\}\}\\right\)\\ \+h\\left\(\\frac\{m\_\{c\}\(\\mathbf\{x\}\)\-\\bar\{m\}\_\{c\}\}\{s\_\{c\}^\{m\}\}\\right\),\(3\)whered¯c\\bar\{d\}\_\{c\}andscds\_\{c\}^\{d\}are the robust location and scale of distances for clean guidance inputs predicted as classcc\. The corresponding robust location and scale of the margins arem¯c\\bar\{m\}\_\{c\}andscms\_\{c\}^\{m\}\. This standardization prevents either component from dominating the other one due its numerical range\. The two standardized terms are then added with equal weight so that hard\-example generation penalizes both poor target\-class proximity and poor separation from competing classes; standardization makes this unweighted combination comparable across the two quantities\.
The verifier is initialized as a copy of the trained base classifier, so both models begin with the same clean decision structure\. The base classifier is then kept fixed, while only the verifier is updated\. To avoid adding target\-sampling randomness to verifier training, a target labelr∈𝒞∖\{y\}r\\in\\mathcal\{C\}\\setminus\\\{y\\\}is selected deterministically during each update\. Letℬ∞\(𝐱,ε\)=\{𝐮:∥𝐮−𝐱∥∞≤ε\}\\mathcal\{B\}\_\{\\infty\}\(\\mathbf\{x\},\\varepsilon\)=\\\{\\mathbf\{u\}:\\lVert\\mathbf\{u\}\-\\mathbf\{x\}\\rVert\_\{\\infty\}\\leq\\varepsilon\\\}denote the permitted perturbation region\. Pixel values are also clipped to the valid input range\. A hard training example𝐱~\\widetilde\{\\mathbf\{x\}\}is approximated by
𝐱~≈argmin𝐮∈ℬ∞\(𝐱,ε\)\[ℓce\(f𝜽\(𝐮\),r\)−λvℓce\(gϕ\(𝐮\),y\)\+λsG\(𝐮,r\)\],\\widetilde\{\\mathbf\{x\}\}\\approx\\underset\{\\mathbf\{u\}\\in\\mathcal\{B\}\_\{\\infty\}\(\\mathbf\{x\},\\varepsilon\)\}\{\\arg\\min\}\\Big\[\\ell\_\{\\mathrm\{ce\}\}\(f\_\{\\boldsymbol\{\\theta\}\}\(\\mathbf\{u\}\),r\)\-\\lambda\_\{\\mathrm\{v\}\}\\ell\_\{\\mathrm\{ce\}\}\(g\_\{\\boldsymbol\{\\phi\}\}\(\\mathbf\{u\}\),y\)\+\\lambda\_\{\\mathrm\{s\}\}G\(\\mathbf\{u\},r\)\\Big\],\(4\)whereℓce\\ell\_\{\\mathrm\{ce\}\}denotes standard cross\-entropy loss\. The coefficientλv\\lambda\_\{\\mathrm\{v\}\}controls the pressure placed on the current verifier\. The coefficientλs\\lambda\_\{\\mathrm\{s\}\}determines the strength of signature camouflage\.
The first term moves the base classifier toward the target label\. The negative second term searches for inputs on which the verifier loses the true label\. The last term encourages the adversarial representation to resemble a clean member of the target class\. We use anL∞L\_\{\\infty\}radius ofε=8/255\\varepsilon=8/255\. The attack performs seven PGD iterations with a step size of2/2552/255\. The verifier coefficient is set toλv=1\\lambda\_\{\\mathrm\{v\}\}=1\. The signature coefficient is set toλs=0\.05\\lambda\_\{\\mathrm\{s\}\}=0\.05\.
The verifier learns to recover the original label from these hard examples while retaining its clean predictive behavior\. Let𝔼ver\\mathbb\{E\}\_\{\\mathrm\{ver\}\}denote expectation over labeled samples from𝒟ver\\mathcal\{D\}\_\{\\mathrm\{ver\}\}\. The verifier objective is
ℒver\(ϕ\)=\\displaystyle\\mathcal\{L\}\_\{\\mathrm\{ver\}\}\(\\boldsymbol\{\\phi\}\)=\{\}𝔼ver\[ℓls\(gϕ\(𝐱\),y\)\+αℓls\(gϕ\(𝐱~\),y\)\\displaystyle\\mathbb\{E\}\_\{\\mathrm\{ver\}\}\\Big\[\\ell\_\{\\mathrm\{ls\}\}\(g\_\{\\boldsymbol\{\\phi\}\}\(\\mathbf\{x\}\),y\)\+\\alpha\\ell\_\{\\mathrm\{ls\}\}\(g\_\{\\boldsymbol\{\\phi\}\}\(\\widetilde\{\\mathbf\{x\}\}\),y\)\(5\)\+βDKL\(π𝜽\(⋅∣𝐱\)∥πϕ\(⋅∣𝐱\)\)\]\.\\displaystyle\+\\beta D\_\{\\mathrm\{KL\}\}\\Big\(\\pi\_\{\\boldsymbol\{\\theta\}\}\(\\cdot\\mid\\mathbf\{x\}\)\\\|\\pi\_\{\\boldsymbol\{\\phi\}\}\(\\cdot\\mid\\mathbf\{x\}\)\\Big\)\\Big\]\.
Here,ℓls\\ell\_\{\\mathrm\{ls\}\}denotes cross\-entropy with label smoothing of 0\.05\. Adversarial recovery is weighted byα=1\.5\\alpha=1\.5\. Clean distillation is weighted byβ=0\.5\\beta=0\.5, andDKLD\_\{\\mathrm\{KL\}\}denotes KL divergence\.
The parameters𝜽\\boldsymbol\{\\theta\}remain fixed during the verifier training\. After every epoch, the verifier is evaluated on𝒟dev\\mathcal\{D\}\_\{\\mathrm\{dev\}\}\. The retained checkpoint maximizes the average of clean accuracy and robust accuracy under a direct joint attack\.
### III\-BPrediction\-Conditioned Verification
The signature bank and activation extractor are removed during inference\. The base classifier first outputs the predicted class
y^\(𝐱\)=argmaxc∈𝒞,f𝜽,c\(𝐱\),\\widehat\{y\}\(\\mathbf\{x\}\)=\\underset\{c\\in\\mathcal\{C\}\}\{\\arg\\max\},f\_\{\\boldsymbol\{\\theta\},c\}\(\\mathbf\{x\}\),\(6\)wheref𝜽,c\(𝐱\)f\_\{\\boldsymbol\{\\theta\},c\}\(\\mathbf\{x\}\)is the logit assigned to classcc\. The detector then measures how strongly the verifier supports this prediction\. Its anomaly score is
A\(𝐱\)=1−πϕ\(y^\(𝐱\)∣𝐱\)\.A\(\\mathbf\{x\}\)=1\-\\pi\_\{\\boldsymbol\{\\phi\}\}\\left\(\\widehat\{y\}\(\\mathbf\{x\}\)\\mid\\mathbf\{x\}\\right\)\.\(7\)
A high value ofA\(𝐱\)A\(\\mathbf\{x\}\)indicates that the verifier assigns little probability to the class selected by the base classifier\. This confidence\-sensitive score is more informative than checking only whether the models produce different top\-1 labels\.
A clean calibration input is denoted by𝐳\\mathbf\{z\}\. For classcc, let𝒟cal\(c\)\\mathcal\{D\}\_\{\\mathrm\{cal\}\}\(c\)contain the calibration inputs whose base prediction iscc\. The class\-specific threshold is
τc=Quantile0\.95\(A\(𝐳\)∣𝐳∈𝒟cal\(c\)\)\.\\tau\_\{c\}=\\operatorname\{Quantile\}\_\{0\.95\}\\left\(A\(\\mathbf\{z\}\)\\mid\\mathbf\{z\}\\in\\mathcal\{D\}\_\{\\mathrm\{cal\}\}\(c\)\\right\)\.\(8\)
If fewer than 20 inputs belong to𝒟cal\(c\)\\mathcal\{D\}\_\{\\mathrm\{cal\}\}\(c\), the global 95th\-percentile clean threshold is used instead\.
The final rejection decision is
R\(𝐱\)=𝕀\[A\(𝐱\)≥τy^\(𝐱\)\],R\(\\mathbf\{x\}\)=\\mathbb\{I\}\\left\[A\(\\mathbf\{x\}\)\\geq\\tau\_\{\\widehat\{y\}\(\\mathbf\{x\}\)\}\\right\],\(9\)whereR\(𝐱\)=1R\(\\mathbf\{x\}\)=1indicates that the input is rejected\. Class\-conditioned calibration accommodates differences among predicted classes while maintaining an empirical clean false\-positive rate near 5%\. Deployment requires one forward pass through each model and does not depend on activation extraction or signature comparison\.
## IVResults
This section presents the experimental protocol and evaluates clean performance, robustness to adaptive and unseen attacks, baseline comparisons, and component ablations\.
### IV\-AExperimental Protocol
TABLE I:Clean performance across three seeds\. Values are reported as mean±\\pmstandard deviation\.MetricCIFAR\-10SVHNSTL\-10Official base acc\. \(%\)92\.98±0\.3292\.98\{\\pm\}0\.3296\.74±0\.1096\.74\{\\pm\}0\.1074\.26±2\.4074\.26\{\\pm\}2\.40Balanced base acc\. \(%\)93\.43±0\.4893\.43\{\\pm\}0\.4896\.13±0\.4596\.13\{\\pm\}0\.4573\.37±2\.3373\.37\{\\pm\}2\.33Verifier acc\. \(%\)85\.70±1\.0685\.70\{\\pm\}1\.0693\.17±0\.2393\.17\{\\pm\}0\.2368\.30±5\.2068\.30\{\\pm\}5\.20Clean FPR \(%\)5\.55±0\.805\.55\{\\pm\}0\.805\.13±0\.335\.13\{\\pm\}0\.335\.32±0\.935\.32\{\\pm\}0\.93Latency \(ms/sample\)0\.359±0\.0090\.359\{\\pm\}0\.0090\.375±0\.1310\.375\{\\pm\}0\.1311\.405±0\.0051\.405\{\\pm\}0\.005TABLE II:Detector performance under direct adaptive attacks\. All entries are percentages reported as mean±\\pmstandard deviation over three seeds\. PGD is seen during the verifier training\. The remaining attacks are unseen\.DatasetAttackAUC↑\\uparrowTPR1%\\mathrm\{TPR\}\_\{1\\%\}↑\\uparrowTPR5%\\mathrm\{TPR\}\_\{5\\%\}↑\\uparrowBypass↓\\downarrowCIFAR\-10PGD79\.10±2\.2279\.10\{\\pm\}2\.2221\.53±2\.7221\.53\{\\pm\}2\.7238\.83±5\.5838\.83\{\\pm\}5\.5857\.23±3\.0457\.23\{\\pm\}3\.04CIFAR\-10APGD79\.06±2\.2179\.06\{\\pm\}2\.2121\.53±2\.4621\.53\{\\pm\}2\.4638\.57±5\.3038\.57\{\\pm\}5\.3057\.50±2\.9557\.50\{\\pm\}2\.95CIFAR\-10AA\+Refine79\.16±2\.1979\.16\{\\pm\}2\.1921\.53±2\.6021\.53\{\\pm\}2\.6039\.00±5\.4339\.00\{\\pm\}5\.4357\.10±2\.9357\.10\{\\pm\}2\.93CIFAR\-10CWL2L\_\{2\}64\.54±1\.3964\.54\{\\pm\}1\.391\.43±0\.351\.43\{\\pm\}0\.352\.73±0\.422\.73\{\\pm\}0\.4295\.63±1\.1795\.63\{\\pm\}1\.17SVHNPGD80\.80±0\.9580\.80\{\\pm\}0\.9524\.80±0\.5324\.80\{\\pm\}0\.5341\.73±2\.6641\.73\{\\pm\}2\.6656\.90±1\.4256\.90\{\\pm\}1\.42SVHNAPGD80\.66±0\.8980\.66\{\\pm\}0\.8924\.80±0\.3024\.80\{\\pm\}0\.3042\.13±1\.7042\.13\{\\pm\}1\.7056\.80±1\.2856\.80\{\\pm\}1\.28SVHNAA\+Refine81\.17±0\.8081\.17\{\\pm\}0\.8025\.83±0\.1225\.83\{\\pm\}0\.1243\.00±2\.0143\.00\{\\pm\}2\.0155\.97±1\.5555\.97\{\\pm\}1\.55SVHNCWL2L\_\{2\}71\.51±1\.6971\.51\{\\pm\}1\.691\.60±0\.101\.60\{\\pm\}0\.102\.37±0\.062\.37\{\\pm\}0\.0694\.53±1\.1594\.53\{\\pm\}1\.15STL\-10PGD79\.86±1\.7479\.86\{\\pm\}1\.7428\.97±3\.4128\.97\{\\pm\}3\.4146\.90±1\.7746\.90\{\\pm\}1\.7751\.30±3\.5651\.30\{\\pm\}3\.56STL\-10APGD79\.92±1\.7179\.92\{\\pm\}1\.7129\.40±3\.4129\.40\{\\pm\}3\.4147\.03±1\.6647\.03\{\\pm\}1\.6651\.20±3\.5851\.20\{\\pm\}3\.58STL\-10AA\+Refine79\.87±1\.7479\.87\{\\pm\}1\.7428\.87±3\.4128\.87\{\\pm\}3\.4146\.87±1\.7246\.87\{\\pm\}1\.7251\.33±3\.5051\.33\{\\pm\}3\.50STL\-10CWL2L\_\{2\}87\.87±1\.7087\.87\{\\pm\}1\.7047\.23±1\.3147\.23\{\\pm\}1\.3162\.00±2\.3562\.00\{\\pm\}2\.3537\.60±1\.6137\.60\{\\pm\}1\.61The main experiments use ResNet\-18 and seeds 0, 1, and 2\. An additional CIFAR\-10 experiment uses DenseNet121 with seed 0\. Main results are reported as the mean and sample standard deviation across three seeds\. The signature soft\-minimum temperature isγ=0\.25\\gamma=0\.25, and eligible layers contain more than 1,000 parameters\. The classifier and verifier use stochastic gradient descent \(SGD\) for 50 and 20 epochs at initial learning rates of 0\.1 and 0\.01, respectively, with momentum 0\.9, weight decay5×10−45\\times 10^\{\-4\}, Nesterov momentum, and cosine decay\. Batch sizes are reduced when needed to fit GPU memory\. Each adaptive evaluation retains 1,000 successful targeted adversarial examples, with 100 examples assigned to each target class\. Only test inputs classified correctly before the attack are eligible\.
PGD, APGD, and the AutoAttack candidate refinement use anL∞L\_\{\\infty\}limit of8/2558/255\. The CW\-style attack uses anL2L\_\{2\}limit of 3\.0\. PGD is the only attack family used during the verifier training\. Detector\-aware PGD and APGD use 80 iterations; CW uses 160 iterations at a learning rate of 0\.01\. AA\+Refine uses 80 detector\-aware PGD refinement iterations\. The additional directL2L\_\{2\}PGD evaluation uses radius 3\.0\. APGD, CW, and AutoAttack candidate refinement are unseen during the training session\.
We report area under the receiver operating characteristic curve, true\-positive rate at 1% and 5% false\-positive rates, and detector bypass rate\. The bypass rate is the fraction of successful targeted adversarial examples accepted by the detector\. For every retained adversarial example, we verified target\-class success and compliance with the specifiedL∞L\_\{\\infty\}orL2L\_\{2\}constraint\. Across the evaluated configurations, targeted attack success ranged from 96\.4% to approximately 100%, with no norm violations\.
Table[I](https://arxiv.org/html/2611.25271#S4.T1)summarizes clean behavior\. Official base accuracy is measured over the complete test set, while balanced base and verifier accuracies use the same class\-balanced set of 2,000 clean test inputs used for detector evaluation\. The calibrated rejection rule yields a clean FPR between 5\.13% and 5\.55%\.
### IV\-BAdaptive Detection and Unseen Attacks
Table[II](https://arxiv.org/html/2611.25271#S4.T2)reports similar detection performance under PGD, APGD, and AA\+Refine across all three datasets\. These results cover both the PGD procedure used during verifier training and two additionalL∞L\_\{\\infty\}evaluation procedures\. At the calibrated 5% false\-positive rate \(FPR\), the detector identifies a substantial portion of successful attacks, with the highest detection rate on STL\-10\.
The CW\-styleL2L\_\{2\}evaluation produces a different dataset\-dependent pattern, with stronger detection on STL\-10 than on CIFAR\-10 or SVHN\. Across the evaluated configurations, targeted attack success is high and all retained examples satisfy their perturbation constraints\. The implications of these differences across attack procedures and perturbation norms are discussed in Section[V](https://arxiv.org/html/2611.25271#S5)\.
### IV\-CBaseline Comparison and Ablation
Each baseline is evaluated using a direct adaptive attack optimized against its own anomaly score\. Table[III](https://arxiv.org/html/2611.25271#S4.T3)compares these results with the proposed detector under direct joint PGD\.
TABLE III:Detector\-awareL∞L\_\{\\infty\}attack results \(mean±\\pmstandard deviation, percent\)\. Each attack targets the corresponding detector score\.DatasetDetectorAUC↑\\uparrowTPR5%\\mathrm\{TPR\}\_\{5\\%\}↑\\uparrowBypass↓\\downarrowCIFAR\-10Proposed79\.10±2\.22\\mathbf\{79\.10\{\\pm\}2\.22\}38\.83±5\.58\\mathbf\{38\.83\{\\pm\}5\.58\}57\.23±3\.04\\mathbf\{57\.23\{\\pm\}3\.04\}CIFAR\-10Activation logistic3\.03±1\.153\.03\{\\pm\}1\.150\.00±0\.000\.00\{\\pm\}0\.0098\.00±0\.8298\.00\{\\pm\}0\.82CIFAR\-10Diagonal Mahalanobis0\.55±0\.110\.55\{\\pm\}0\.110\.00±0\.000\.00\{\\pm\}0\.00100\.00±0\.00100\.00\{\\pm\}0\.00CIFAR\-10Energy1\.92±1\.241\.92\{\\pm\}1\.240\.60±0\.610\.60\{\\pm\}0\.6199\.27±0\.6899\.27\{\\pm\}0\.68CIFAR\-10MSP2\.72±1\.282\.72\{\\pm\}1\.280\.50±0\.530\.50\{\\pm\}0\.5399\.37±0\.6199\.37\{\\pm\}0\.61SVHNProposed80\.80±0\.95\\mathbf\{80\.80\{\\pm\}0\.95\}41\.73±2\.66\\mathbf\{41\.73\{\\pm\}2\.66\}56\.90±1\.42\\mathbf\{56\.90\{\\pm\}1\.42\}SVHNActivation logistic8\.26±1\.338\.26\{\\pm\}1\.330\.00±0\.000\.00\{\\pm\}0\.0098\.60±0\.5698\.60\{\\pm\}0\.56SVHNDiagonal Mahalanobis16\.17±0\.5616\.17\{\\pm\}0\.561\.43±0\.381\.43\{\\pm\}0\.3896\.97±0\.8696\.97\{\\pm\}0\.86SVHNEnergy7\.00±0\.157\.00\{\\pm\}0\.152\.37±0\.322\.37\{\\pm\}0\.3297\.00±0\.3597\.00\{\\pm\}0\.35SVHNMSP9\.10±1\.109\.10\{\\pm\}1\.102\.27±0\.512\.27\{\\pm\}0\.5197\.17±0\.9597\.17\{\\pm\}0\.95STL\-10Proposed79\.86±1\.74\\mathbf\{79\.86\{\\pm\}1\.74\}46\.90±1\.77\\mathbf\{46\.90\{\\pm\}1\.77\}51\.30±3\.56\\mathbf\{51\.30\{\\pm\}3\.56\}STL\-10Activation logistic36\.96±2\.8536\.96\{\\pm\}2\.850\.00±0\.000\.00\{\\pm\}0\.00100\.00±0\.00100\.00\{\\pm\}0\.00STL\-10Diagonal Mahalanobis11\.83±2\.6511\.83\{\\pm\}2\.650\.17±0\.150\.17\{\\pm\}0\.1599\.70±0\.1799\.70\{\\pm\}0\.17STL\-10Energy1\.16±1\.221\.16\{\\pm\}1\.220\.07±0\.120\.07\{\\pm\}0\.1299\.93±0\.1299\.93\{\\pm\}0\.12STL\-10MSP1\.63±1\.441\.63\{\\pm\}1\.440\.03±0\.060\.03\{\\pm\}0\.0699\.90±0\.1799\.90\{\\pm\}0\.17
\(a\)Adversarial detection rate across adaptive attacks when the detection threshold is calibrated to incorrectly reject 5% of clean inputs\.\(b\)Detector bypass under direct adaptive attacks\.
Fig\. 2:Adaptive detection results across CIFAR\-10, SVHN, and STL\-10\. The verifier behaves similarly across the threeL∞L\_\{\\infty\}attacks, while CWL2L\_\{2\}varies across datasets\. It also achieves substantially lower bypass rates than the baseline detectors\.The proposed verifier achieves the strongest direct adaptive detection results across the three datasets and reported metrics in Table[III](https://arxiv.org/html/2611.25271#S4.T3)\. Its AUC exceeds that of the strongest baseline on every dataset, while the directly attacked confidence\-based and feature\-based detectors have substantially higher bypass rates\. The comparison isolates each detector’s behavior when the attack is optimized against its own score\.
TABLE IV:Direct full versus no\-signature comparison under each detector\-aware attack\. Values are mean±\\pmsample standard deviation in percent, over three seeds\. Each evaluation retains 1,000 successful targeted examples\.DatasetAttackFull AUC↑\\uparrowNo\-signature AUC↑\\uparrowFull bypass↓\\downarrowNo\-signature bypass↓\\downarrowCIFAR\-10PGD79\.10±2\.2279\.10\{\\pm\}2\.2283\.24±0\.5083\.24\{\\pm\}0\.5057\.23±3\.0457\.23\{\\pm\}3\.0451\.07±2\.3051\.07\{\\pm\}2\.30CIFAR\-10APGD79\.06±2\.2179\.06\{\\pm\}2\.2183\.33±0\.4583\.33\{\\pm\}0\.4557\.50±2\.9557\.50\{\\pm\}2\.9551\.00±2\.3651\.00\{\\pm\}2\.36CIFAR\-10AA\+Refine79\.16±2\.1979\.16\{\\pm\}2\.1983\.38±0\.5083\.38\{\\pm\}0\.5057\.10±2\.9357\.10\{\\pm\}2\.9350\.77±2\.1750\.77\{\\pm\}2\.17CIFAR\-10CWL2L\_\{2\}64\.54±1\.3964\.54\{\\pm\}1\.3965\.66±0\.1565\.66\{\\pm\}0\.1595\.63±1\.1795\.63\{\\pm\}1\.1795\.37±0\.6195\.37\{\\pm\}0\.61SVHNPGD80\.50±1\.1480\.50\{\\pm\}1\.1485\.54±0\.5085\.54\{\\pm\}0\.5056\.35±1\.4856\.35\{\\pm\}1\.4853\.15±2\.1953\.15\{\\pm\}2\.19SVHNAPGD80\.35±1\.0280\.35\{\\pm\}1\.0285\.49±0\.5385\.49\{\\pm\}0\.5356\.10±0\.5756\.10\{\\pm\}0\.5753\.25±2\.4753\.25\{\\pm\}2\.47SVHNAA\+Refine80\.98±1\.0380\.98\{\\pm\}1\.0385\.75±0\.4285\.75\{\\pm\}0\.4255\.20±1\.1355\.20\{\\pm\}1\.1352\.30±1\.5652\.30\{\\pm\}1\.56SVHNCWL2L\_\{2\}70\.54±0\.0870\.54\{\\pm\}0\.0873\.79±2\.0373\.79\{\\pm\}2\.0395\.10±0\.8595\.10\{\\pm\}0\.8595\.20±0\.0095\.20\{\\pm\}0\.00Table[IV](https://arxiv.org/html/2611.25271#S4.T4)provides paired comparisons of the full verifier and a no\-signature control\. Both variants use the same verifier training, distillation, calibration, and rejection rule; the control setsλs=0\\lambda\_\{\\mathrm\{s\}\}=0instead of 0\.05\. Signature guidance improves clean verifier accuracy on CIFAR\-10 and, to a smaller extent, SVHN\. The no\-signature variant obtains higher adaptive AUC across the paired attacks\. These configurations also provide a two\-point comparison ofλs\\lambda\_\{\\mathrm\{s\}\}at fixedγ=0\.25\\gamma=0\.25\.
The CIFAR\-10 component ablation further separates the contributions of the training objectives\. Removing adaptive training reduces PGD AUC from 81\.66% to 0\.98% and increases bypass from 53\.80% to 99\.80%\. Removing distillation preserves comparable AUC but reduces clean verifier accuracy\.
The additional DenseNet121 experiment on CIFAR\-10 exhibits the same clean\-accuracy pattern as ResNet\-18: signature guidance improves clean verifier accuracy, while the no\-signature variant obtains higher adaptive AUC\. The matched evaluation uses the same source images and target classes for both variants, with each attack optimized against its own verifier\. A separate, larger PGD evaluation of the full DenseNet121 model provides an additional check\.
## VDiscussion and Future Directions
The experiments suggest that effective verification benefits from separating the classifier’s prediction from the evidence used to accept it\. Detector\-aware training teaches the verifier to respond to examples designed to challenge its rejection rule, rather than relying solely on the confidence or activation score of the fixed classifier\. The consistent behavior across PGD, APGD, and AA\+Refine further suggests that the learned verification signal captures properties shared by differentL∞L\_\{\\infty\}attack procedures\. This motivates further study of prediction\-conditioned verification as a modular complement to existing classifiers\.
The ablations also distinguish the objectives served by adaptive training, signature guidance, and distillation\. Adaptive training is the primary source of the observed detection benefit, whereas signature guidance improves the verifier’s clean predictions in the evaluated comparisons\. The difference between the full and no\-signature configurations suggests that clean predictive fidelity and adaptive detection should be considered jointly when choosing the signature weight\. A natural extension is to investigate adaptive weighting schedules or class\-dependent signature guidance, rather than using a single fixed coefficient\. Distillation can be studied within the same framework as a way to preserve the fixed classifier’s clean decision structure while the verifier learns from difficult examples\.
The prediction\-conditioned rejection rule provides a separate point of control\. Class\-specific thresholds allow the detector’s operating point to reflect differences among predicted classes without requiring activation extraction at inference\. Future work could examine alternative calibration objectives, including class\-wise detection requirements and selective\-classification criteria, to determine how the verifier’s score can be used in applications with different acceptance policies\. The training\-only signature bank also keeps deployment independent of signature storage and comparison, making inference\-cost studies and integration with existing classification pipelines relevant next steps\.
Finally, the DenseNet121 experiment provides an initial indication that the clean\-accuracy effect of signature guidance is not specific to ResNet\-18\. Broader studies could extend the paired evaluation to additional architectures, datasets, and training seeds\. The evaluation with multiple attack objectives and perturbation norms also motivates a systematic study of multi\-norm and multi\-attack training curricula\. Such evaluations would help characterize which aspects of the learned verification signal transfer across attack objectives and perturbation models\.
## Acknowledgment
AI\-use declaration:OpenAI ChatGPT assisted with drafting and editing the Abstract and Sections I–V\. The authors take full responsibility for the manuscript\.
## References
- \[1\]N\. Carlini and D\. Wagner\(2017\)Towards Evaluating the Robustness of Neural Networks\.In2017 IEEE Symposium on Security and Privacy \(SP\),Vol\.,pp\. 39–57\.External Links:ISSN 2375\-1207,[Document](https://dx.doi.org/10.1109/SP.2017.49)Cited by:[§I](https://arxiv.org/html/2611.25271#S1.p1.1)\.
- \[2\]A\. Madry, A\. Makelov, L\. Schmidt, D\. Tsipras, and A\. Vladu\(2017\)Towards deep learning models resistant to adversarial attacks\.arXiv preprint arXiv:1706\.06083\.Cited by:[§I](https://arxiv.org/html/2611.25271#S1.p1.1),[§II\-A](https://arxiv.org/html/2611.25271#S2.SS1.p1.1)\.
- \[3\]N\. Carlini and D\. Wagner\(2017\)Adversarial examples are not easily detected: bypassing ten detection methods\.InProceedings of the 10th ACM Workshop on Artificial Intelligence and Security,AISec ’17,New York, NY, USA,pp\. 3–14\.External Links:ISBN 9781450352024,[Document](https://dx.doi.org/10.1145/3128572.3140444)Cited by:[§I](https://arxiv.org/html/2611.25271#S1.p1.1),[§II\-D](https://arxiv.org/html/2611.25271#S2.SS4.p1.1)\.
- \[4\]F\. Tramer, N\. Carlini, W\. Brendel, and A\. Madry\(2020\)On adaptive attacks to adversarial example defenses\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 1633–1645\.Cited by:[§I](https://arxiv.org/html/2611.25271#S1.p1.1),[§II\-D](https://arxiv.org/html/2611.25271#S2.SS4.p1.1)\.
- \[5\]D\. Hendrycks and K\. Gimpel\(2017\)A baseline for detecting misclassified and out\-of\-distribution examples in neural networks\.InInternational Conference on Learning Representations,Cited by:[§I](https://arxiv.org/html/2611.25271#S1.p2.1),[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p2.1)\.
- \[6\]W\. Liu, X\. Wang, J\. Owens, and Y\. Li\(2020\)Energy\-based out\-of\-distribution detection\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 21464–21475\.Cited by:[§I](https://arxiv.org/html/2611.25271#S1.p2.1),[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p2.1)\.
- \[7\]K\. Lee, K\. Lee, H\. Lee, and J\. Shin\(2018\)A simple unified framework for detecting out\-of\-distribution samples and adversarial attacks\.InAdvances in Neural Information Processing Systems,Vol\.31\.Cited by:[§I](https://arxiv.org/html/2611.25271#S1.p2.1),[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[8\]J\. H\. Metzen, T\. Genewein, V\. Fischer, and B\. Bischoff\(2017\)On detecting adversarial perturbations\.InInternational Conference on Learning Representations,Cited by:[§I](https://arxiv.org/html/2611.25271#S1.p2.1),[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[9\]N\. Palumbo, Y\. Guo, X\. Wu, J\. Chen, Y\. Liang, and S\. Jha\(2023\)Two heads are actually better than one: towards better adversarial robustness via transduction and rejection\.arXiv preprint arXiv:2305\.17528\.Cited by:[§I](https://arxiv.org/html/2611.25271#S1.p2.1),[§II\-C](https://arxiv.org/html/2611.25271#S2.SS3.p1.1)\.
- \[10\]F\. Croce and M\. Hein\(2020\)Reliable evaluation of adversarial robustness with an ensemble of diverse parameter\-free attacks\.InProceedings of the 37th International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.119,pp\. 2206–2216\.Cited by:[§I](https://arxiv.org/html/2611.25271#S1.p5.1),[§II\-D](https://arxiv.org/html/2611.25271#S2.SS4.p1.1)\.
- \[11\]A\. Vashagh, R\. Razavi\-Far, M\. Meymani, and B\. Biggio\(2026\)Recent advances in adversarial attacks on model utility, privacy, and explainability: a comprehensive survey\.TechRxiv2026\(0305\)\.External Links:[Document](https://dx.doi.org/10.36227/techrxiv.177272853.30003431/v1)Cited by:[§II](https://arxiv.org/html/2611.25271#S2.p1.1)\.
- \[12\]M\. Meymani, R\. Razavi\-Far, A\. Vashagh, and B\. Biggio\(2026\)Defense against adversarial attacks: foundations, strategies, and future directions\.Preprints\.External Links:[Document](https://dx.doi.org/10.20944/preprints202607.1022.v1)Cited by:[§II](https://arxiv.org/html/2611.25271#S2.p1.1)\.
- \[13\]S\. Dai, C\. Cianfarani, V\. Sehwag, P\. Mittal, and A\. Bhagoji\(2025\)Adapting to evolving adversaries with regularized continual robust training\.InProceedings of the 42nd International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.267,pp\. 11954–12000\.Cited by:[§II\-A](https://arxiv.org/html/2611.25271#S2.SS1.p1.1)\.
- \[14\]R\. Razavi\-Far, E\. Hallaji, A\. Fathalizadeh, M\. Wu, and M\. Rostami\(2026\)Towards self\-adaptive learning: a comprehensive survey on continual learning under harsh conditions\.Neurocomputing698,pp\. 134136\.External Links:ISSN 0925\-2312,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.neucom.2026.134136)Cited by:[§II\-A](https://arxiv.org/html/2611.25271#S2.SS1.p1.1)\.
- \[15\]F\. Wang, X\. Zuo, H\. Huang, G\. Chen, and H\. Qian\(2026\)SeRI: gradient\-free sensitive region identification in decision\-based black\-box attacks\.InInternational Conference on Learning Representations,Vol\.2026,pp\. 123055–123083\.Cited by:[§II\-A](https://arxiv.org/html/2611.25271#S2.SS1.p1.1)\.
- \[16\]T\. Pang, C\. Du, Y\. Dong, and J\. Zhu\(2018\)Towards robust detection of adversarial examples\.InAdvances in Neural Information Processing Systems,Vol\.31\.Cited by:[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[17\]S\. Speakman, S\. Sridharan, S\. Remy, K\. Weldemariam, and E\. McFowland\(2018\)Subset scanning over neural network activations\.arXiv preprint arXiv:1810\.08676\.Cited by:[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[18\]H\. Fisher, M\. Shahar, and Y\. S\. Resheff\(2024\)Neural fingerprints for adversarial attack detection\.arXiv preprint arXiv:2411\.04533\.Cited by:[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[19\]Y\. Li, P\. Angelov, and N\. Suri\(2023\)Fuzzy detectors against adversarial attacks\.In2023 IEEE Symposium Series on Computational Intelligence \(SSCI\),pp\. 306–311\.External Links:[Document](https://dx.doi.org/10.1109/SSCI52147.2023.10372061)Cited by:[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[20\]F\. Mumcu and Y\. Yilmaz\(2025\)Universal and efficient detection of adversarial data through nonuniform impact on network layers\.Transactions on Machine Learning Research\.Note:External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=0CY5APFnFI)Cited by:[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[21\]Z\. Chen, R\. Wang, J\. Xiang, Y\. Yu, X\. Xia, S\. Ji, Q\. Xuan, and X\. Yang\(2024\)GGT: graph\-guided testing for adversarial sample detection of deep neural network\.Computers & Security140,pp\. 103710\.External Links:ISSN 0167\-4048,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.cose.2024.103710)Cited by:[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[22\]H\. Mirzaei Sadeghlou, M\. Nafez, J\. Habibi, M\. Sabokrou, and M\. H\. Rohban\(2025\)Adversarially robust anomaly detection through spurious negative pair mitigation\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 32113–32140\.Cited by:[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[23\]G\. Dong, R\. Bresson, and H\. Boström\(2025\)Detect adversarial examples with exchangeability martingale\.InProceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications,Proceedings of Machine Learning Research, Vol\.266,pp\. 758–761\.Cited by:[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[24\]A\. Vashagh and R\. Razavi\-Far\(2026\)GraphRectify: graph\-based transfer of adversarial example detectors across neural networks\.arXiv preprint arXiv:2610\.10423\.Cited by:[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p1.1)\.
- \[25\]X\. Yin, C\. Zhang, J\. Steele, N\. Shavit, and T\. Wang\(2026\)Scalable energy\-based models via adversarial training: unifying discrimination and generation\.InInternational Conference on Learning Representations,Vol\.2026,pp\. 68385–68424\.Cited by:[§II\-B](https://arxiv.org/html/2611.25271#S2.SS2.p2.1)\.
- \[26\]A\. K\. Nirala, J\. Tian, O\. Fakorede, and M\. Atsague\(2025\)AlignFix: fixing adversarial perturbations by agreement checking for adversarial robustness against black\-box attacks\.Transactions on Machine Learning Research\.Note:External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=XgK05fssnx)Cited by:[§II\-C](https://arxiv.org/html/2611.25271#S2.SS3.p1.1)\.
- \[27\]V\. Shah, T\. Chaudhari, and N\. Manwani\(2024\)Towards calibrated losses for adversarial robust reject option classification\.arXiv preprint arXiv:2410\.10736\.Cited by:[§II\-C](https://arxiv.org/html/2611.25271#S2.SS3.p1.1)\.
- \[28\]A\. Heng and H\. Soh\(2026\)Know when to abstain: optimal selective classification with likelihood ratios\.InInternational Conference on Learning Representations,Vol\.2026,pp\. 37816–37839\.Cited by:[§II\-C](https://arxiv.org/html/2611.25271#S2.SS3.p1.1)\.
- \[29\]W\. Liu, W\. Zhang, K\. Yang, Y\. Chen, K\. Guo, and J\. Wei\(2024\)Enhancing generalization in few\-shot learning for detecting unknown adversarial examples\.Neural Processing Letters56\(2\),pp\. 85\.Cited by:[§II\-C](https://arxiv.org/html/2611.25271#S2.SS3.p2.1)\.
- \[30\]H\. Lee, S\. Cho, and C\. Kim\(2025\)Indirect gradient matching for adversarial robust distillation\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 70007–70028\.Cited by:[§II\-C](https://arxiv.org/html/2611.25271#S2.SS3.p2.1)\.
- \[31\]Z\. Li, L\. Lu, X\. Zheng, S\. Liang, T\. Chen, Y\. Zhou, and S\. Pang\(2026\)FERD: fairness\-enhanced data\-free adversarial robustness distillation\.InInternational Conference on Learning Representations,Vol\.2026,pp\. 152021–152043\.Cited by:[§II\-C](https://arxiv.org/html/2611.25271#S2.SS3.p2.1)\.
- \[32\]A\. Athalye, N\. Carlini, and D\. Wagner\(2018\)Obfuscated gradients give a false sense of security: circumventing defenses to adversarial examples\.InProceedings of the 35th International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.80,pp\. 274–283\.Cited by:[§II\-D](https://arxiv.org/html/2611.25271#S2.SS4.p1.1)\.
- \[33\]M\. Pintor, L\. Demetrio, A\. Sotgiu, A\. Demontis, N\. Carlini, B\. Biggio, and F\. Roli\(2022\)Indicators of attack failure: debugging and improving optimization of adversarial examples\.InAdvances in Neural Information Processing Systems,Vol\.35,pp\. 23063–23076\.External Links:[Document](https://dx.doi.org/10.52202/068431-1676)Cited by:[§II\-D](https://arxiv.org/html/2611.25271#S2.SS4.p1.1)\.
- \[34\]F\. Tramer\(2022\)Detecting adversarial examples is \(Nearly\) as hard as classifying them\.InProceedings of the 39th International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.162,pp\. 21692–21702\.Cited by:[§II\-D](https://arxiv.org/html/2611.25271#S2.SS4.p1.1)\.
- \[35\]F\. Croce, S\. Gowal, T\. Brunner, E\. Shelhamer, M\. Hein, and T\. Cemgil\(2022\)Evaluating the adversarial robustness of adaptive test\-time defenses\.InProceedings of the 39th International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.162,pp\. 4421–4435\.Cited by:[§II\-D](https://arxiv.org/html/2611.25271#S2.SS4.p1.1)\.Similar Articles
Robustness Meets Uncertainty: Evidential Adversarial Training for Robust Selective Classification
This paper introduces Evidential Adversarial Training (EV-AT), a method that improves the robustness-uncertainty trade-off in classifiers by combining an evidence-based loss with robust evidence alignment, achieving state-of-the-art results on selective classification benchmarks.
@omarsar0: NEW AI paper worth bookmarking. This is something I called early, and this paper confirms it: verification has emerged …
This paper from Stanford, NVIDIA, and UC Berkeley introduces LLM-as-a-Verifier, a training-free verification framework that uses continuous scoring from LLM logits to improve accuracy across coding, robotics, and medical domains, achieving state-of-the-art results on multiple benchmarks.
Towards Adversarially Robust Multigraph Neural Networks for Financial Crime Analysis
Delft University of Technology researchers present the first scalable framework for structural adversarial attacks and adversarial training on multigraph neural networks used for anti-money laundering detection, introducing a cycle-based regularizer to keep perturbations plausible and showing that adversarial training restores detection robustness.
Verification-Aware Training for Speculative Decoding
Verification-Aware Training (VAT) improves draft models for speculative decoding by simulating sequential verification during training and adapting loss weights to acceptance patterns, leading to enhanced acceptance length and inference speedup.
Rethinking the Transferable Adversarial Attacks and Robust Defense in Federated Learning
This paper analyzes the transferability of adversarial attacks in federated learning systems and proposes a defense mechanism based on adversarial training to enhance model robustness.