Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face Imagery

TL;DR

Proto-LeakNet tackles the challenge of attributing synthetic face imagery by exploiting signal-leak traces embedded in diffusion latents. It introduces a latent-domain framework with temporal attention and a prototype-based attribution head, complemented by density-based open-set evaluation to identify unseen generators without retraining. The method achieves a Macro AUC of 98.13% on closed-set data, remains robust under aggressive post-processing, and reveals interpretable latent geometry through prototypes and feature-wise gates. Representation-level analysis via KDE demonstrates separability between known and unseen generators, validating open-set generalization without supervision. This work provides a scalable, interpretable approach for provenance attribution in synthetic media.

Abstract

The growing sophistication of synthetic image and deepfake generation models has turned source attribution and authenticity verification into a critical challenge for modern computer vision systems. Recent studies suggest that diffusion pipelines unintentionally imprint persistent statistical traces, known as signal-leaks, within their outputs, particularly in latent representations. Building on this observation, we propose Proto-LeakNet, a signal-leak-aware and interpretable attribution framework that integrates closed-set classification with a density-based open-set evaluation on the learned embeddings, enabling analysis of unseen generators without retraining. Acting in the latent domain of diffusion models, our method re-simulates partial forward diffusion to expose residual generator-specific cues. A temporal attention encoder aggregates multi-step latent features, while a feature-weighted prototype head structures the embedding space and enables transparent attribution. Trained solely on closed data and achieving a Macro AUC of 98.13%, Proto-LeakNet learns a latent geometry that remains robust under post-processing, surpassing state-of-the-art methods, and achieves strong separability both between real images and known generators, and between known and unseen ones. The codebase will be available after acceptance.

Figures (4)

• Figure 1: Proto-LeakNet: given an input image $x$, latent features are extracted from the pretrained Stable Diffusion 2.1 Variational Autoencoder (VAE) in Block 1 (Latent Feature Construction). For each diffusion step $t\!\in\!\{0,5,10\}$, we extract a $z_t$ latent which gets normalized with dimension $(4, 32, 32)$. In Block 2 (Discriminative Backbone), each $z_t$ is encoded by a ResNet-18, producing embeddings $\{h_t\}$ that are temporally aggregated through the Attention Temporal Pooling module to yield a single representation $\bar{h}$. Block 3 (Prototype-Based Attribution) computes distances between $\bar{h}$ and class prototypes $p_{c,m}$, modulated by a feature-wise gating vector $w$ obtained from a small MLP. The resulting attention-weighted distances are aggregated via a LogSumExp scoring function to produce class probabilities $\pi_c$. Symbols: "+" denotes the weighted sum over attention coefficients across timesteps, and "*" indicates the element-wise product between the encoder output and the gating vector.
• Figure 2: Interpretability of Proto-LeakNet (t-SNE Evolution). Layer-wise evolution of the embedding space, illustrating an approximate progression of how prototypes and attention progressively refine and generalize class separation across layers.
• Figure 3: Histogram of the Top-1 accuracy distributions per class for each method, from raw to step 1-3.
• Figure 4: Impact of attention configurations on latent-space separation. (a) Disabling attention for both domains leads to overlapping closed and open clusters. (b) Enabling attention for both aligns open embeddings with the closed distribution. (c) The asymmetric setup, attention active only for closed samples, produces well-separated clusters. (d) Applying the same setup as (c) with real CELEBA-HQ images, separability is preserved.