Analysis of an AI based code distinguisher

What it claims

Quentin Lavit, Julien Seinturier and Pascal Véron (LORIA and Université de Toulon) argue that DeepDistinguisher does not learn the algebraic structure of Goppa codes. DeepDistinguisher is a transformer-like model that separates Goppa codes from random linear codes. Its accuracy instead traces to general combinatorial properties of the generator matrix. The paper appeared at the 20th IMA International Conference on Cryptography and Coding and was posted to ePrint on 20 October 2026.

How it works

The analysis has three stages:

1. Artifact check. Saliency maps and attention visualisation expose low-weight rows in the random training instances, which the model partly uses as a cue.
2. Artifact removal. With that artifact removed, the distinguisher still reaches significant accuracy. The low-weight-row cue alone does not explain the model, so the authors look further.
3. XOR rule. A hand-built rule applies pairwise XOR between rows of the generator matrix. It is motivated by the constraints that the minimum distance of Goppa codes imposes on row combinations. The rule's predictions are compared with the model's.

The final test uses random matrices built to satisfy both the minimum-distance constraint and the XOR constraint.

Evidence

DeepDistinguisher's predictions correlate highly with the XOR rule. The correlation is strongest for high-rate codes, a qualifier the catalog summary omits. On the constrained random matrices, the neural distinguisher loses most of its predictive power. The abstract gives no accuracy figures, correlation coefficients, code parameters or dataset sizes.

Limits

Only the abstract and metadata were available, so every claim here is qualitative. The abstract does not say how much accuracy remains after the artifact is removed, or how weak the XOR correlation becomes at lower rates. It also does not say which Goppa parameters were tested. Before reusing the training sets or the conclusions, check the full paper for those numbers. The paper is licensed CC BY-NC, which rules out commercial reuse of the text and figures.