@gbinj95
🚨 @gensynai "Hail to the Thief"
Decentralized GRPO is vulnerable: Few malicious nodes can poison LLMs via high-reward sequences → full model drift in <50 rounds.
Attacks:
- In-context (alter math/code)
- Out-of-context (irrelevant text)
→ 100% success on math/code
Defenses:
1. Log-prob verification (homogeneous)
2. LLM-as-a-Judge (heterogeneous)
→ Block 100% attacks, minimal overhead
First systematic study of dRL security.
📄 arxiv.org/abs/2511.09780
🔗 gensyn.ai/blog/hail-to-t…
#DecentralizedAI #GRPO #LLMSecurity #Gensyn
0 reply
0 recast
0 reaction