Adversarial attacks are usually seen as a threat to AI reliability. In this preprint, Illia Horenko reframes them as an optimisation problem: find the smallest perturbation of an input that produces the largest change in a model's output. He argues that an entropy-based perspective makes this problem tractable and yields provable conditions for when an attack is globally optimal and unique, using criteria that are cheap to verify.
Building on that, the author introduces 'self-attacking' learning algorithms. Instead of pitting two networks against each other as in GANs, the algorithm uses the mathematical optimality criterion to generate new training points on its own. This replaces one of the trained agents with a direct mathematical rule, which the author says allows the model to learn from far fewer data points.
Tests span synthetic problems and real-world cases from hydrodynamics and biomedicine. According to the paper, the resulting models are orders of magnitude simpler and cheaper, yet achieve better predictive performance than a wide set of contemporary tools—including boosted random forests, deep neural networks, and transformer-based foundation models—while using only a small fraction of the training data.
The claims are strong, but they come from a single author's comparisons against the tools he chose to include. As a preprint, the work has not yet undergone peer review, and independent benchmarks would be needed to confirm the reported advantages.