Research map / Image generation / Autoregressive and latent generation

BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Generalizable fake image detection · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BIAS-ID, a framework that quantifies transformation biases in AI-generated image detectors via score shifts and an aggregated transform sensitivity (ATS) metric, and applies it to six detectors across two datasets.

Problem

AI-generated image detectors may rely on spurious correlations from transformation biases in training data (e.g., compression, resizing, rotation), but there is no established protocol to measure how biased a detector is, as distinct from its lack of robustness.

Approach

Measure bias from the detector's raw predicted scores rather than accuracy: compute per-image score shifts under transformations at multiple levels, average them over real and fake sets, and aggregate into an ATS score whose absolute sum over classes flags systematic class-directed bias.

Key results

• B-Free and RINE are remarkably unbiased w.r.t. JPEG compression, while AIDE and SPAI show significant score shifts.
• DRCT and RINE are strongly biased w.r.t. WebP compression, in opposite directions.
• Bicubic resizing yields slightly higher average ATS(X_R, X_F) (0.167) than bilinear (0.132), averaged over detectors and datasets.

Surprising findings

• Strong rotation effects (-90 and 90 degrees) shift real image scores as well, not only fake ones, contrary to the explanation of prior work.
• Detection performance can decline without a systematic class bias, and bias effects are not always visible in AUC/performance metrics alone.

Limitations

• Analysis of five biases and two datasets is not exhaustive; unknown dataset biases may confound results.
• Framework cannot assess content/semantic biases since it relies on transforming images and comparing scores.
• Synthbuster and SynthCLIC results deviate for some detector/bias combinations.

Experimental setup

Datasets: RAISE-1k Synthbuster SynthCLIC CLIC2020 ImageNet LSUN COCO FLUX subset

Benchmarks: Synthbuster SynthCLIC RAISE-1k

Metrics: mean score shift aggregated transform sensitivity (ATS) AUC accuracy

Baselines: UnivFD DRCT RINE AIDE SPAI B-Free

Builds on: UnivFD DRCT RINE AIDE SPAI B-Free

Closest papers by meaning

• A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.90)
• Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.89)
• Semi-Truths: A Large-Scale Dataset of AI-Augmented Images for Evaluating Robustness of AI-Generated Image detectors (NeurIPS 2024 · similarity 0.87)
• Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.87)
• CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.87)
• Manifold Induced Biases for Zero-shot and Few-shot Detection of Generated Images (ICLR 2025 · similarity 0.87)