La Belle Noiseuse: Adversarial Noise between Ichnography and Semantic Topology

Adversarial perturbations — subtle modifications to images that fool neural networks into categorical misrecognition — reveal something fundamental about the relationship between noise and form in computational regimes of vision. The paper situates this contemporary phenomenon of weaponised noise within a literary and art-historical precursor: in Balzac's The Unknown Masterpiece (1831), the painter Frenhofer unveils what he believes to be a figurative masterpiece only to confront his spectators with a ‘painted wall’, a chaos of colours and indecipherable lines from which only a foot emerges. Michel Serres (1983) interprets Frenhofer's canvas as ichnography: white noise as the constitutive ground of all possible representational profiles, the formless matrix from which form emerges. Giorgio Agamben (1994) reads the scene as an irreducible doubling, in which the work splits between the artist's lived reality and the spectator's aesthetic judgement, a perceptual rift that anticipates the incommensurability between sensory and statistical perception at the heart of machine vision.

Adversarial perturbations, I argue, restage Frenhofer's scene with new urgency. Neural networks process images within hyperdimensional vector spaces, carving statistical classification boundaries that constitute what I call a ‘stochastic gestalt’: form perceived not in the iconic surface of the image but in the mathematical distributions that demarcate categorical identity. Adversarial perturbations exploit the topology of these boundaries, pushing images beyond their classifications through barely perceptible noise. When this noise is considered aesthetically, it coalesces into quasi-figuration — variegated textures, proto-landscapes — arriving at the juncture of noise and form that Serres theorises through Frenhofer's canvas, yet under radically different conditions.

Serres's ontologising of noise as originary plenitude, as the primordial sea from which form is born, however cannot account for the material-technical conditions that produce adversarial noise. This noise does not precede classification; it is generated by the classificatory apparatus itself, a collateral of machine learning’s iterative refinement of decision boundaries through backpropagation. I therefore propose a materialist reframing of Serres: adversarial perturbations reveal noise not as the a priori fount of representation but as the constitutive remainder of computational identification, what the system produces in the act of subsuming visual complexity under classificatory logics. To read a landscape in this noise is not to witness form emerging from the formless, as Serres's ichnography would suggest, but to perform a second-order figuration upon the collateral of statistical perception. Such a reading folds the representational back into the operational, and in doing so complicates Serres in a further sense: if ichnography is the set of all possible profiles, then the figurative patterns legible within adversarial noise are not drawn from that originary set but produced after the fact, phantom profiles generated by the very classificatory violence they seem to precede. It is this recursive complicity of noise and form, rather than their sequential unfolding, that adversarial perturbations make newly legible.