The Implicit Bias of Structured State Space Models Can Be Poisoned With Clean Labels
This reveals a vulnerability in SSMs, an increasingly popular alternative to transformers, highlighting a critical need for research into mitigating clean-label poisoning attacks.
The paper demonstrates that structured state space models (SSMs), which rely on implicit bias for generalization, can be made to fail generalization through clean-label poisoning, where specially crafted training examples with correct labels distort the implicit bias.
Neural networks are powered by an implicit bias: a tendency of gradient descent to fit training data in a way that generalizes to unseen data. A recent class of neural network models gaining increasing popularity is structured state space models (SSMs), regarded as an efficient alternative to transformers. Prior work argued that the implicit bias of SSMs leads to generalization in a setting where data is generated by a low dimensional teacher. In this paper, we revisit the latter setting, and formally establish a phenomenon entirely undetected by prior work on the implicit bias of SSMs. Namely, we prove that while implicit bias leads to generalization under many choices of training data, there exist special examples whose inclusion in training completely distorts the implicit bias, to a point where generalization fails. This failure occurs despite the special training examples being labeled by the teacher, i.e. having clean labels! We empirically demonstrate the phenomenon, with SSMs trained independently and as part of non-linear neural networks. In the area of adversarial machine learning, disrupting generalization with cleanly labeled training examples is known as clean-label poisoning. Given the proliferation of SSMs, we believe that delineating their susceptibility to clean-label poisoning, and developing methods for overcoming this susceptibility, are critical research directions to pursue.