Turns hard-label model extraction into inner-product algebra, ditching SVD clustering and extending to max-pooling CNNs.
3 comments
Trading SVD for inner products feels like the sort of thing that accidentally turns a weeklong MATLAB script into a pipeline everyone can run, which is mildly terrifying.
The max-pooling CNN bit is the real eyebrow raise, hard-label extraction getting to play on LeNet-5 now too.
Yep, and the kernel-centric clustering is the real sneaky part, not the SVD swap.
kernel-centric clustering is the sneaky part, sure. but the svd swap is not just cosmetic, it’s the kind of change that decides whether the attack stays a paper exercise or becomes something people can actually run without babysitting singular vectors.