The field of AI research is moving towards addressing critical challenges in safety and security. Recent developments focus on improving the robustness of large language models, detecting data contamination, and enhancing privacy protections. Innovative methods, such as fine-grained iterative adversarial attacks and semantically-aware privacy agents, are being proposed to tackle these challenges. Noteworthy papers in this area include those that introduce novel frameworks for detecting data contamination, propose adaptive defense strategies against harmful fine-tuning, and develop test-time debiasing methods for vision-language models. These advancements have significant implications for the development of more secure and trustworthy AI systems.
Advances in AI Safety and Security
Sources
Fine-Grained Iterative Adversarial Attacks with Limited Computation Budget
Semantically-Aware LLM Agent to Enhance Privacy in Conversational AI Services
Detecting Data Contamination in LLMs via In-Context Learning
Contrastive Knowledge Transfer and Robust Optimization for Secure Alignment of Large Language Models
Adaptive Defense against Harmful Fine-Tuning for Large Language Models via Bayesian Data Scheduler
Un-Attributability: Computing Novelty From Retrieval & Semantic Similarity
Visual Backdoor Attacks on MLLM Embodied Decision Making via Contrastive Trigger Learning
Best Practices for Biorisk Evaluations on Open-Weight Bio-Foundation Models
Probing Knowledge Holes in Unlearned LLMs
EL-MIA: Quantifying Membership Inference Risks of Sensitive Entities in LLMs
SegDebias: Test-Time Bias Mitigation for ViT-Based CLIP via Segmentation
Black-Box Membership Inference Attack for LVLMs via Prior Knowledge-Calibrated Memory Probing
Optimizing AI Agent Attacks With Synthetic Data
Contamination Detection for VLMs using Multi-Modal Semantic Perturbation
PrivacyCD: Hierarchical Unlearning for Protecting Student Privacy in Cognitive Diagnosis
PETRA: Pretrained Evolutionary Transformer for SARS-CoV-2 Mutation Prediction
REMIND: Input Loss Landscapes Reveal Residual Memorization in Post-Unlearning LLMs
Reusing Pre-Training Data at Test Time is a Compute Multiplier
Benchmark Designers Should "Train on the Test Set" to Expose Exploitable Non-Visual Shortcuts
Forgetting is Everywhere