Tag
The paper proposes a lightweight defense framework to enhance the empirical privacy of DP-SGD without theoretical privacy cost, validated through extensive audits across models, datasets, and threat models.
RAGuard is a layered defense framework for Retrieval-Augmented Generation (RAG) systems that uses adversarial fine-tuning of the retriever and a label-free filter (ZKIP) to achieve zero attack success against corpus poisoning, maintaining high retrieval accuracy.
This paper presents MIPIAD, a multilingual defense framework against indirect prompt injection attacks using a hybrid of Qwen2.5-based classifiers and TF-IDF features with meta-ensemble learning. It demonstrates strong performance on English and Bangla benchmarks, achieving high F1 and AUROC scores while reducing cross-lingual gaps.