Yevgeniy Vorobeychik

A handful of verified poisoned examples, not just clean ones, dramatically boosts data poisoning defense against clean-label attacks.

Malicious instruction detection can be significantly improved by adapting adversarial training to the context of the task, leading to better robustness against evolving attack strategies.

Real-time user feedback can now directly optimize diffusion models for personalized recommendations, eliminating the need for extensive preference data.

IMPFM achieves unprecedented global exploration in feedback-driven search by leveraging multi-particle interactions to prevent mode collapse and reward over-optimization.

A malicious parameter server can exploit parameter-efficient fine-tuning to implant a privacy backdoor that reconstructs training data with alarming accuracy.

Stop training in isolation: LNTrust lets decentralized models learn *who* to trust during training, so they can collaborate effectively at deployment, boosting accuracy and cutting communication costs.

Adversarial perturbations in LLMs have an exploitable low-rank structure, enabling more efficient and effective black-box attacks.

Multi-agent RL can find optimal strategies to defend vehicular routing networks against false data injection attacks, significantly outperforming traditional baselines.

Key contribution not extracted.

Deter unauthorized LLM distillation by subtly rewriting reasoning traces, degrading student model performance while actually *improving* the teacher's.