Hush-Up attack halts distributed LLM inference
Original title: The Blind Spots of Detection: Adversarial Perturbation Attacks against Watchdogs in Distributed LLM Inference
A single malicious worker forces premature LLM termination with small output perturbations that evade local activation-based verification.
Hush-Up is a stealthy availability attack on decentralized LLM inference. A malicious worker profiles hardware errors and optimizes adversarial perturbations on activations to maximize EOS token probability. This forces generation halt while evading watchdogs that tolerate benign discrepancies. Existing defenses fail to detect the sabotage or misattribute it.
- Forces premature LLM generation termination
- Evades activation-based verification
- Perturbations too small to indict locally
- Breaks distributed inference globally
Impact: halts LLM generation while evading detection
- Products
- distributed LLM inference
- AI models
- LLM
From the original article
1 min read at arxiv.orgDecentralizing large language model (LLM) inference across untrusted Web nodes requires robust verification to ensure computational integrity. Existing activation-based verification and detection mechanisms validate intermediate outputs against reference values, permitting bounded numerical deviations to accommodate benign cross-hardware discrepancies.
By Siu Wai Shi · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.