cs.LGNov 9, 2026

When Are Predictions Enough? An Evaluation Protocol for Frozen Expert Composition

Authors: Alessandro Pereira, Keith J Ransom, Lewis Mitchell

Organizations: Australian Institute for Machine Learning (AIML), Adelaide University · School of Computer Science and Information Technology, Adelaide University · Adelaide Data Science Centre (ADSC), School of Mathematical Sciences, Adelaide University

Abstract

Practitioners often combine independently trained models while keeping their weights frozen. These experts are commonly combined through their final predictions, although their internal features may retain useful information that those predictions omit. We ask when combining features improves on combining predictions, and how to make that comparison fair. We introduce a reusable evaluation protocol, apply it to 21 frozen deepfake detectors from DF40 (Yan et al., 2024), and audit how evaluation choices change the conclusions. The protocol uses strong prediction baselines, compares simple and larger feature combiners, varies architecture and training data separately, and reports results by held-out generator with train/test overlap controls. Under leave-one-generator-out evaluation, logistic regression on the features of six architecture-diverse experts improves macro AUROC by 0.081 over the best tested prediction combiner in each fold, with gains on seven of eight generators. The larger network shows no clear improvement over the linear model (approximately 3.5M versus 10K parameters), suggesting that access to richer information matters more than additional combiner capacity on this pool. Features may help when they retain information that the predictions omit and that transfers to new generators. Predictions may be enough when the strongest prediction baseline already performs well, while neither approach reliably compensates for poor expert transfer. Weakening prediction baselines on the overlap-affected standard split inflates the apparent feature advantage by up to 0.172 AUROC. Overlap-aware evaluation also changes the estimated gain, though the contrast does not isolate overlap. Results by generator and expert pool show where the advantage holds, weakens, or reverses. Without these controls, the same benchmark can support different conclusions.

Explore similar work

Nov 11, 2026

Honest Fusion Evaluation for Cross-Domain Deepfake Detection: Leakage Audits and Tail Robustness

Cross-domain video deepfake detection increasingly uses frozen foundation-model experts, combined by calibrated score averaging, a per-video router picking one expert, or a learned head over concatenated features. Expressive routing and feature fusion are often reported to beat plain averaging. Across a 22-dataset frozen-expert testbed, we find those gains are largely evaluation artifacts. A corner theorem locates the gap routing aims to close: because AUC depends only on score ordering, the best any score-space router can do is an oracle reading the test label to pick each video's expert, a ceiling no realisable, label-blind router reaches. The theorem does not order real routers against a calibrated global average; the case against routing is empirical, and no router we test beats that average. Feature fusion reads raw features, falls outside the theorem, and is audited empirically. Reported gains trace to leakage: holding out one dataset leaks siblings sharing identities and reals (over-crediting routing), while holding out one manipulation keeps that dataset's reals in training (over-crediting feature fusion). A 2×22\times 2 over fold and evaluation granularity pins the +2.6+2.6pp feature-fusion "win" to the fold alone: 12/12 significant under leaky folds vs. 0/12 under honest family-held-out folds, where both gains vanish. Fusion still pays off in the tail of hard videos and families: as experts are added, videos every expert misclassifies drop from 15.2% to 2.3%, and global calibrated fusion (InvB) beats the best single expert on worst-family AUC, selective risk (AURC), and low-false-positive recall, stably across protocols. Practitioners should calibrate and combine globally, hold out whole dataset families, and measure tail performance. For score-space routing under ranking metrics, honest evaluation and expert diversity matter more than combiner complexity.
Sep 29, 2025

R2^2M: Real-Aware Residual Model Merging for Robust and Generalizable Deepfake Detection

Deepfake generators evolve rapidly, making exhaustive data collection and repeated retraining impractical. Unlike generic multi-task settings, deepfake specialists share a common binary objective (Real vs. Fake) and mainly differ in generator-specific artifacts. However, naive parameter arithmetic can induce unintended decision-boundary shifts, causing unstable ranking behavior and degraded AUC under domain shift. We propose R2M, a training-free merging framework that decomposes task vectors into a shared component and generator-specific residuals, linking parameter-space updates to logit-space behavior. Offline spectral construction identifies shared and residual subspaces, and online routing selects residuals via first-order gradient-residual alignment, predicting per-sample logit updates under linearization. R2M is both efficient and interpretable: expensive computations are performed offline, while online inference requires only a single forward-backward pass with lightweight inner-product routing. The same formulation enables diagnostic analysis through margin and routing statistics. Experiments demonstrate consistently strong performance across in-domain, cross-domain, and unseen settings, highlighting R2M as an interpretable and scalable approach to training-free deepfake model merging.
Feb 23, 2026

Pixels Don't Lie (But Your Detector Might): Bootstrapping MLLM-as-a-Judge for Trustworthy Deepfake Detection and Reasoning Supervision

Deepfake detection models often generate natural-language explanations, yet their reasoning is frequently ungrounded in visual evidence, limiting reliability. Existing evaluations measure classification accuracy but overlook reasoning fidelity. We propose DeepfakeJudge, a framework for scalable reasoning supervision and evaluation, that integrates an out-of-distribution benchmark containing recent generative and editing forgeries, a human-annotated subset with visual reasoning labels, and a suite of evaluation models, that specialize in evaluating reasoning rationales without the need for explicit ground truth reasoning rationales. The Judge is optimized through a bootstrapped generator-evaluator process that scales human feedback into structured reasoning supervision and supports both pointwise and pairwise evaluation. On the proposed meta-evaluation benchmark, our reasoning-bootstrapped model achieves an accuracy of 96.2%, outperforming 30x larger baselines. The reasoning judge attains very high correlation with human ratings and 98.9% percent pairwise agreement on the human-annotated meta-evaluation subset. These results establish reasoning fidelity as a quantifiable dimension of deepfake detection and demonstrate scalable supervision for interpretable deepfake reasoning. Our user study shows that participants preferred the reasonings generated by our framework 70% of the time, in terms of faithfulness, groundedness, and usefulness, compared to those produced by other models and datasets. All of our datasets, models, and codebase are open-sourced.