The Future of AI Cryptanalysis

Theoretical Edge of AI in Cryptanalysis

Matthew Green argues that over-engineering by cryptographers provides a margin of safety against AI-driven attacks. He notes that human cryptanalysis has historically advanced through jumps, such as differential and side-channel attacks, not just brute force. He also mentions the concern about implementation issues in cryptographic systems, which can lead to side channels and vulnerabilities.

Implementation and Side Channels

Cryptographic systems are more than just algorithms. The implementations designed by cryptographers and put up on the NIST Competition Web Site are not always what one might hope from the theoretical strength of the algorithms. For example, the AES implementation, designed for speed, opened up a series of side channels. These side channels can leak secrets to the line, which can then be passed through upstream routers into the cloud. This leakage is a concern, especially for devices like smart meters and IoT devices, where users cannot see the leakage but the NSA can.

IoT Devices and Side Channels

The concern about side channels extends to IoT devices. Many IoT devices have been designed with side channels in mind, which can be exploited by AI models. The concern is that even post-quantum cryptography (PQC) implementations might have side channels, making them vulnerable to attacks. This is a significant issue, as IoT devices are becoming more ubiquitous and are often used in critical systems.

Over-Engineering and Resilience

Over-engineering in many subsystems can lead to a stronger and more resilient system as a whole. This is a theoretical concept that Green believes can provide a margin of safety against AI-driven attacks. However, the practical implementation of this concept is still uncertain and requires further research.