CASCADE 2026, Regensburg. To appear in Springer LNCS, volume 16532.
A Multi-head CNN-based Side-channel Attack on the Energy-efficient DIZY Stream Cipher
When a chip encrypts data, its power consumption leaks a little about the secret key. I trained convolutional neural networks on power traces captured from an FPGA running DIZY, a lightweight stream cipher built for IoT devices, and used them to recover parts of the key.
DIZY is harder to attack than the usual AES benchmark: it runs 32 S-boxes in the same clock cycle, which buries each one in noise, and each S-box output maps to only two key bits. A multi-head network with a shared backbone, predicting the full 160-bit internal state at once, converged faster and recovered more key bits than separate networks per S-box group.
- First deep-learning side-channel attack on a small-state stream cipher like DIZY
- Up to 11 of 32 two-bit key groups recovered in the upper 64 key bits
- As few as 514 traces needed on average, from datasets of up to 500,000 traces
Read the preprint See the code
I presented the work at the CASCADE 2026 conference in Regensburg. It grew out of my master's thesis at the Barkhausen Institut.