arXiv: Image Encryption Algorithm Based on Convolutional Neural Networks and Dynamic S-Box Generation

This publication from arXiv presents a novel image encryption algorithm that integrates convolutional neural networks with dynamic S-box generation. While not a regulatory change itself, it signals a significant technical development in AI-driven cryptography that may impact compliance under the EU AI Safety framework. The paper demonstrates how neural networks can be used to create adaptive encryption methods, potentially challenging existing standards for data protection and AI system transparency. This is relevant because the EU AI Act and related safety regulations require that AI systems, including those used for security, be explainable, robust, and free from hidden vulnerabilities.

Organizations in sectors handling sensitive visual data—such as healthcare, finance, defense, and critical infrastructure—are most affected. Compliance teams in these sectors must assess whether their current encryption methods rely on static or predictable algorithms that could be rendered obsolete by such dynamic approaches. Additionally, any deployment of AI-based encryption would need to meet the EU’s requirements for high-risk AI systems, including risk management, human oversight, and documentation.

Compliance teams should immediately review their organization’s encryption and data protection policies to identify any reliance on static S-box or non-adaptive methods. They should also initiate a gap analysis against the AI Safety framework’s transparency and robustness requirements, particularly for any AI systems used in security contexts. Finally, they should monitor the European Commission’s guidance on AI-based cryptography and prepare to update internal standards if this technology becomes commercially deployed.