adversarial-robustness

This paper from Carnegie Mellon researchers shows that giving an LLM judge more compute doesn't fix oversight failures when it must check many requirements in one call. It proposes sharding—dividing requirements into smaller groups handled by separate calls—which improves accuracy, resists presentation-based adversarial attacks, and can make a weaker sharded judge match a more capable holistic judge.

This paper empirically investigates how pruning, adversarial training, and hardware-induced weight faults jointly affect the reliability of convolutional neural networks, finding that adversarial training increases sensitivity to stuck-at-zero faults while pruning has little effect on fault sensitivity.

This paper proposes a domain-knowledge-free metacognitive layer for fusing multiple pre-trained ViT-based perception models, using label vector pools and consistency-based abduction. It matches majority-vote baselines on clean data and is particularly robust against coordinated label-flipping attacks.

A systematic cross-architecture empirical study measuring the trustworthiness cost of domain adaptation in small language models, finding that safety-preserving fine-tuning strategies do not reliably transfer alignment.

This paper proposes efficient search methods to locate verdict boundaries in Branch and Bound (BaB) neural network verification, leveraging path monotonicity to skip irrelevant subproblems and improve verification efficiency.

This paper assesses the adversarial robustness of five Arabic language models under character, word, and sentence-level attacks, showing that diacritic insertion can reduce accuracy by 92% and adversarial training improves resilience but has limitations.

This paper introduces a source-aware reranking method for RAG that incorporates domain-informed source reliability priors, improving Precision@5 from 0.48 to 0.72 on a 120-document health corpus and reducing adversarial document retrieval.

This paper proposes HarmAlign, a method that applies function-preserving spectral deformation along an estimated contrastive activation subspace to block harmful fine-tuning of open-weight models while preserving benign adaptability, with finite-sample guarantees and empirical validation.

This paper introduces a novel training criterion to reduce reliance on shortcuts in automated spoken English proficiency assessment systems, aiming to prevent test-takers from inflating scores without genuine improvement. Experiments on audio and text-based systems show reduced correlation with exploitable features, bringing automated scores closer to human judgment.

This paper challenges the prevailing view that rote memorization causes training data exposure to reconstruction attacks, showing instead that adversarial non-robust features are the true cause. The authors introduce AntiAdversarial Training (AT-AT) that intentionally learns non-robust features to achieve superior reconstruction defense and higher accuracy.

This paper presents a rigorous theoretical framework for adversarial robustness in multilayered perceptrons by reducing the problem to lattice traversal, introducing both sound and complete interval certifications with formal guarantees.

This paper investigates formal mechanisms, such as Mediation, to maintain market stability among self-interested LLM agents (DeepSeek-V3) in a simulated marketplace, finding that Mediation enables recovery even under sustained adversarial attacks.

Introduces LipSSD, a Lipschitz-constrained variant of Single Shot MultiBox Detector that improves adversarial robustness in object detection while being attack-agnostic and complementary to adversarial training. Evaluated on Pascal VOC, LARD, and KITTI datasets.

This paper studies adversarial robustness in programming-by-example systems, where an adversary corrupts input-output examples to mislead the synthesizer. It introduces version-space partition aggregation (VPA) as a defense and evaluates it across multiple benchmarks, finding that low-margin tasks are vulnerable and VPA helps only when partition vote margins are preserved.

This paper investigates the cross-domain generalization failure of lightweight ML models for IIoT intrusion detection, finding they rely on coarse port features and that adversarial robustness does not correlate with cross-network performance.

This paper introduces a meta-learning framework for anytime-valid certified robustness that uses sequential E-processes to adaptively allocate compute, achieving a 20-fold reduction in sample complexity compared to traditional randomized smoothing while maintaining rigorous statistical guarantees.

Yuvion LLM is a large language model designed for adversarial robustness and content safety, achieving state-of-the-art performance on safety benchmarks and outperforming larger models such as GPT-5.4 and Qwen3-MAX.

This paper evaluates the reliability of automated judges used to measure attack success rates (ASR) in LLM jailbreak research, finding that both safety classifiers and LLM-as-judges have significant calibration and adversarial robustness issues that undermine reported ASR numbers.

This paper presents a rigorous N-qubit theory of stochastic quantum neural networks (SQNNs) for adversarially robust network intrusion detection, proving a decoherence-contraction theorem and showing that depolarising noise provides robustness against adversarial attacks, with experiments on the NSL-KDD dataset.

MorphStrata introduces a layer-specific stochastic noise injection strategy for generating diverse student models in a Moving Target Defense framework to enhance adversarial robustness in time-series forecasting, achieving up to 97.97% improvement in RMSE under BIM attacks with minimal training overhead.