- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
OPFL checks a random sample of training steps by repeating the computation through a privacy-preserving protocol. Because honest computations can produce slightly different results on different systems, it first measures the expected differences and uses these as a tolerance for accepting updates. The contribution is a combination of private checking, numerical tolerances, and penalties for detected cheating. Tests on three models found no successful attacks or incorrect rejections in the evaluated settings; a separate cost estimate suggests that checking only a small fraction of steps could substantially reduce verification costs.
Critical strengths
Comparing discrepancies avoids requiring exact agreement between GPU execution and MPC. The combination of various gating mechanisms and statistical constraints are well motivated.
The evaluation is broad, covering three architectures, full training and LoRA, multiple optimizers, dataset changes, sequence-length changes, and a second GPU type. Reporting false rejection alongside attack success is useful. Table 3 also demonstrates reliability.
Critical weaknesses
Every attacked update was checked in the experiments. The proposed low-cost setting checks only 1% but it isn't clear if that would be sufficient to stop an attacker poisoning the model.
I had a look at the supplied code, and I note that it calculates relative differences differently from the paper’s description, and the Qwen attack checker omits one of the checks.
I also note that the speed estimates exclude checking each update against its recorded commitment, which could significantly reduce the apparent time savings when only 1% of updates are replayed.
Questions for authors
Do poisoning attacks need to be consistently applied to be effective? At a 1% checking rate, can't an attacker reasonable perform one (or several) one-shot poisoning attacks without discovery?
Do the attacks still fail if they can choose smaller changes to the submitted gradients, or try to damage the model gradually over several training rounds? In the current tests, how often did the attacker find a modified gradient that passed verification, rather than give up and submit the original honest gradient?
Can you update the table 4 timing table to ablate the runtime implications of the different parts of the proposed approach