Differential-Neural Cryptanalysis of ChaCha and Related ARX Stream Ciphers

What it is

A systematic differential-neural cryptanalysis study of the "Latin Dances" ARX stream cipher family (ChaCha, Salsa, Forró), by Nimai Parsa and Nitin Kumar Sharma. Published as an IACR ePrint preprint (2026/2382), received 2026-10-06, approved 2026-10-08.

How it works

Six neural distinguisher architectures (Gohr's residual network, DBitNet, Inception network, MLP-ResNet, SE-ResNet, self-attention network) are trained and compared on round-reduced ChaCha, in both a classical single-pair setting and a multi-pair setting that combines up to 16 ciphertext pairs via a log-odds rule. The same approach is then applied across Salsa and Forró. A separate experimental variant family, "XOR-1," modifies the cipher to isolate how much of the distinguishability comes specifically from round-1 diffusion.

What to do

The concrete results, beyond what the summary states:

• 4.5-round Salsa distinguisher: accuracy 0.53 with one pair, rising to 0.60 with 16 pairs.
• 1.25-round Forró distinguisher: accuracy of 1.0 (perfect) for both single-pair and multi-pair settings.
• The XOR-1 variant family is offered as a reusable experimental tool for isolating first-round diffusion contribution — useful for anyone extending this analysis to other ARX designs.
• The explainability analysis on 3-round ChaCha identifies that just five round-1 truncated differentials account for 99% of the high-confidence ciphertext pairs the distinguisher relies on — a specific, falsifiable structural claim about where the learned signal actually lives.

Caveats

The Salsa result (0.53–0.60 accuracy) is barely above random guessing (0.50), so its practical cryptanalytic significance is marginal despite being described as "first-ever." The Forró result of perfect accuracy applies only at a very shallow 1.25 rounds. These are preprint results (received Oct 6, approved Oct 8, 2026) and have not gone through peer review.