Test whether the stealthy-flip pool predicts trained outcomes across more class pairs

Determine whether the stealthy-flip pool predicts the trained victim’s attack outcome on more than the four CIFAR-10 source-to-target class pairs evaluated in the paper.

Background

The stealthy-flip pool correctly distinguishes the single tested zero-pool pair from the three tested non-zero-pool pairs, but it does not order the severity of outcomes within the non-zero group. Because only four of the 90 ordered CIFAR-10 class pairs were evaluated with victim training, the authors identify broader validation as an unresolved question. Testing additional pairs would determine whether the pool’s coarse predictive relationship generalizes beyond the limited experiment.

References

Finally, the stealthy-flip pool is computable for all 90 ordered CIFAR-10 pairs in seconds, so testing whether it predicts the trained outcome on more than four of them is limited by victim training time alone, and is worth the compute.

— Space Over Statistic, Retraining Over Ranking: Auditing Training Labels Under Label-Only Backdoors  (2610.13051 - Dey et al., 9 Oct 2026) in Section “Conclusion and future work”