The day in conversation for 2026-11-12

discussion

Transcript

The discussion: Elias: That routing-aware backdoor concept in BadMoE is fascinating because it moves beyond just poisoning a single expert; it targets the sparsity inherent in MoE structures, specifically focusing on those experts with low routing mass. What do you think that implies for how we secure deployment pipelines? Nadia: It’s unsettling, especially since the attack relies on measuring contribution; if they can reliably map which specific components are underutilized, that suggests a new vector for injecting malicious behavior without significantly increasing the model's overall training footprint. I was thinking earlier about our work on adversarial prompting—if you can target a specific expert path, the perturbation required might be much smaller than trying to affect the entire network uniformly. Priya: That leads me to how we verify those paths. I’m more interested in the detection side, specifically what Known By Their Actions is suggesting. The fact that agent actions achieve up to 96% F1 in identifying the underlying LLM implies that behavioral fingerprinting is becoming a viable alternative to traditional static analysis for model provenance. But does that detection hold up when we consider models where the action space itself is highly contextual or dynamic, like those used in multi-agent systems? Elias: Exactly. If we look at the constraints on SnapAudit, it seems like differential privacy verification is getting more formal, moving from just checking noise levels to using simulated noise within snapshots to audit the learning process itself. The detail there is

More episodes

← Home