Restore `margin` parameter to `find_adversarial_example` - Githubissues

vtjeng / MIPVerify.jl

Evaluating Robustness of Neural Networks with Mixed Integer Programming

MIT License

113 stars 31 forks source link

Restore `margin` parameter to `find_adversarial_example` #160

Open vtjeng opened 8 months ago

vtjeng commented 8 months ago

Adding a margin would enable users to ensure that the target category has logits strictly larger than any non-target category, accounting for solver tolerances. For more details, see https://github.com/vtjeng/MIPVerify.jl/issues/158#issuecomment-1997804778.

Sketch

Ensure that margin is applied in all solve modes (closest; worst).
Add tests that fail at HEAD and pass with the new behavior. It would probably be best to use a simple NN.

NOTE: This would partially undo https://github.com/vtjeng/MIPVerify.jl/pull/65.