Beyond Scalar Probes: Exploiting Vector-Valued Outputs in ReLU Networks For Signature Extraction
Vector-valued Jacobian analysis improves cryptanalytic signature extraction from ReLU networks at low precision.
The paper revisits cryptanalytic extraction of ReLU networks by analyzing full vector-valued outputs across adjacent linear regions rather than a single output component. A rank-one characterization of Jacobian differences recovers the usual row-signature information and also exposes column-side structure. Experiments show that structure improves numerical estimation under float64, float32, and float16, extending prior analysis into practical low-precision settings.
- Studies vector outputs across adjacent linear regions.
- Jacobian differences give row and column signature data.
- Estimation improves at float64, float32, and float16.
- Extends extraction analysis into practical low precision.
Full article104 words · extracted from arxiv.org · click to collapse
We revisit cryptanalytic extraction of ReLU networks from a geometric and algebraic perspective. Rather than restricting attention to a single output component, we study the full vector-valued behavior across adjacent linear regions. This leads to a rank-one characterization of Jacobian differences that recovers the usual row-signature information while also revealing complementary column-side information. Our experiments show how this additional structure can be used in ex- traction and improves numerical estimation under different numerical- precision regimes (float64, float32 and float16). We extend the analysis beyond the high-precision and output-rounding settings commonly con- sidered in the literature towards the low-precision settings encountered in many practical settings.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.35487