Decoy language models often cost attackers more energy than defenders
Original title: Make the Adversary Spend More Energy: Generative Deception against Language-Model Attackers in Critical Infrastructure
In 330 sessions, the attacker-to-defender energy ratio stayed above one in every configuration once diagnostic calls were counted.
Researchers measured an autonomous language-model attacker against a decoy that uses a defending model to interpret commands and choose responses. Across 330 sessions, eight model pairs, two attacker profiles, and three counter-injection levels, both sides ran on identical GPUs with energy measured above idle. With thinking disabled, attackers used more energy than defenders in 41 of 48 configurations excluding diagnostic calls, and in all 48 when those calls were included. Enabling thinking raised defender energy 8 to 17 times and reversed one pair.
Unconfirmed: The SCADA deception architecture is proposed from the measurements and is not reported as deployed.
- Token ratios stayed within about a factor of two when generation was not truncated.
- Counter-injection raised attacker tokens about 9% per level without a pooled energy change.
- Authors propose a SCADA decoy with private physical state per identified adversary.
- The proposal uses deterministic protocol and process emulation.
From the original article
1 min read at arxiv.orgLanguage-model attackers consume computational resources as they explore and interact with a target, creating an opportunity for cyber deception to impose costs. Whether such engagements favor the defender requires measuring both sides. We evaluate an autonomous language-model attacker against a decoy that uses a defending language model to analyze attacker commands and select defensive responses.
By Vlad Mihal · 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.