Adversarial Perturbation

This essay examines adversarial perturbations — small, deliberate modifications to images that cause machine learning algorithms to misrecognise their content — as a juncture of aesthetics and computational epistemology. Situating these perturbations within a longer history of tactical image operations, the essay argues that they function as meta-operational images, intervening upon the logic of machine vision rather than within it. Through an analysis of how neural networks process images as points within hyperdimensional classification spaces, the essay proposes the concept of a ‘stochastic gestalt’, a form that emerges not in the visual surface of the image but in the statistical distributions demarcating decision boundaries. Where Shklovsky's 'defamiliarisation' resists the automatisation of human perception, adversarial perturbations reveal that human and machine vision operate in ontologically incommensurable regimes of the visible: the alienation is mutual. Reading Balzac's The Unknown Masterpiece (1831) as both anticipation and allegory, the essay traces how figuration persists at the fold between noise and image, between stochastic and aesthetic.