Wenbo Jiang

E2E speech models are vulnerable to a stealthy DoS attack that can drastically increase their output length and resource consumption without altering the original input.

Malicious modifications to router weights can turn Mixture-of-Experts models into trigger-controlled bottlenecks, revealing a critical vulnerability in AI serving architectures.

Backdoor attacks can be stealthily inherited during model merging, but DiffSafeMerge ensures zero worst-target ASR while preserving image quality across multiple datasets.

InkShield reduces the risk of handwriting forgery by over 90% while keeping the original style visually intact.

Auditing CLIP backdoors reveals that textual encoders can become carriers of adversarial behavior, exposing a critical vulnerability in model deployment.

Existing corpus poisoning attacks falter in realistic RAG systems, but the new CRCP framework ensures adversarial effectiveness by aligning chunking and reranking processes.

By enforcing graph isomorphism across counterfactual inputs, UGID reveals that debiasing LLMs can be achieved by directly manipulating internal representations and attention mechanisms.

Worried about compromised cloud environments skewing your endpoint auditing? vCause offers a verifiable causality analysis system with negligible overhead.

Backdoor attacks can now hide in plain sight: by delaying activation, common words become viable triggers, opening a new, stealthier attack surface in pre-trained models.