OpenNovelty.org

Enabling Your Forensic Detector Know ​How Well​ It Performs on Distorted Samples

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 7.0Download Report PDF
image distortionsforensicsqualityconfidence
Abstract:

Generative AI has substantially facilitated realistic image synthesizing, posing great challenges for reliable forensics. When image forensic detectors are deployed in the wild, the inputs usually undergone various distortions including compression, rescaling, and lossy transmission. Such distortions severely erode forensic traces and make a detector fail silently—returning an over-confident binary prediction while being incapable of making reliable decision, as the detector cannot explicitly perceive the degree of data distortion. This paper argues that reliable forensics must therefore move beyond "is the image real or fake?" to also ask "how trustworthy is the detector's decision on the image?" We formulate this requirement as Detector's Distortion-Aware Confidence (DAC): a sample-level confidence that a given detector could properly handle the input. Taking AI-generated image detection as an example, we empirically discover that detection accuracy drops almost monotonically with full-reference image quality scores as distortion becomes severer, while such references are in fact unavailable at test time. Guided by this observation, the Distortion-Aware Confidence Model (DACOM) is proposed as a useful assistant to the forensic detector. DACOM utilizes full-reference image quality assessment to provide oracle statistical information that labels the detectability of images for training, and integrates intermediate forensic features of the detector, no-reference image quality descriptors and distortion-type cues to estimate DAC. With the estimated confidence score, it is possible to conduct selective abstention and multi-detector routing to improve the overall accuracy of a detection system. Extensive experiments have demonstrated the effectiveness of our approach.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces Distortion-Aware Confidence (DAC) for forensic detectors, specifically targeting AI-generated image detection under real-world distortions like compression and rescaling. It resides in the 'Distortion-Aware Confidence Modeling' leaf, which contains only two papers including this one. This leaf sits within the broader 'Confidence Estimation and Calibration Methods' branch, indicating a relatively sparse but well-defined research direction. The small sibling count suggests this specific formulation—explicitly modeling detector confidence as a function of distortion severity in forensic contexts—occupies a niche position rather than a crowded subfield.

The taxonomy reveals neighboring work in 'Uncertainty-Based Confidence Estimation' (three papers using prediction variance) and 'Semantic Perturbation for Calibration' (one paper on vision-language models), both addressing confidence reliability but without explicit distortion modeling. The 'Detection Architecture Adaptation' branch (thirteen papers across four leaves) focuses on architectural robustness rather than confidence estimation, while 'Robustness Evaluation and Benchmarking' (ten papers) emphasizes systematic testing frameworks. The paper's approach bridges distortion modeling with confidence calibration, distinguishing it from pure robustness studies or general uncertainty quantification methods that do not integrate image quality descriptors.

Among twenty-one candidates examined, none clearly refute the three core contributions: DAC formulation (ten candidates examined, zero refutable), DACOM architecture (ten candidates examined, zero refutable), and the two-stage severity binning pipeline (one candidate examined, zero refutable). The sibling paper in the same leaf likely addresses related distortion-aware confidence but in a different application domain or with alternative modeling strategies. The limited search scope—top-K semantic matches plus citation expansion—means the analysis captures closely related work but may not cover all tangential approaches in adjacent fields like adversarial robustness or domain adaptation.

Given the sparse taxonomy leaf and absence of refuting candidates among twenty-one examined papers, the work appears to occupy a distinct position within distortion-aware confidence modeling. However, the small search scale and narrow leaf population suggest caution: the novelty assessment reflects proximity to known methods rather than exhaustive field coverage. The forensic detection framing and explicit quality-descriptor integration differentiate this work from broader calibration or robustness studies, though the degree of technical novelty relative to the single sibling paper remains unclear without deeper comparison.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
21
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: estimating detection confidence under image distortions. The field addresses how object detectors maintain reliable confidence scores when input images suffer from noise, blur, compression artifacts, or other degradations. The taxonomy reveals several complementary research directions. Confidence Estimation and Calibration Methods focus on modeling and correcting confidence scores to reflect true detection reliability, often incorporating distortion-aware mechanisms (e.g., Forensic Detector Distortion[0], Reliability Prediction Distortion[20]). Detection Architecture Adaptation for Distortions explores specialized network designs that handle geometric or photometric distortions, such as fisheye lenses (OmniDet Fisheye[14], FE-Det Fisheye[16]) or adverse weather (Hazy Rainy Detection[15]). Image Enhancement and Preprocessing for Detection aims to restore degraded inputs before detection, while Robustness Evaluation and Benchmarking provides systematic frameworks for testing detector performance under controlled perturbations (Robustness Verification Perturbations[19], Corrupted Input Robustness[37]). Domain-Specific Detection Applications and Specialized Image Processing Techniques address niche scenarios like aerial imagery (Aerial Robustness[11]) or underground environments (Hybrid Underground Enhancement[5]), and Perceptual and Cognitive Uncertainty Modeling examines how humans and machines represent uncertainty (Objective Uncertainty Evaluation[27]). A particularly active line of work centers on distortion-aware confidence modeling, where methods explicitly account for input quality when predicting detection reliability. Forensic Detector Distortion[0] sits within this branch, emphasizing how forensic analysis scenarios demand accurate confidence under various image manipulations. This contrasts with broader robustness studies like Pretraining Robustness Uncertainty[12], which examine how pretraining strategies influence model resilience, or with calibration-focused approaches such as Evidential Confusion Ignorance[22], which model epistemic uncertainty without explicit distortion modeling. Nearby, Reliability Prediction Distortion[20] shares a similar emphasis on predicting confidence degradation but may differ in application domain or distortion types. Meanwhile, works like Object Verbalized Confidence[1] and VL Uncertainty Hallucination[2] explore confidence estimation in vision-language models, highlighting a shift toward multimodal uncertainty. The original paper's focus on forensic contexts positions it at the intersection of distortion modeling and domain-specific reliability, addressing both technical calibration challenges and the practical need for trustworthy detection under real-world image degradations.

Claimed Contributions

Distortion-Aware Confidence (DAC) formulation for forensic detectors

The authors introduce a formal definition of Distortion-Aware Confidence as a sample-level probability that a forensic detector will correctly classify a given distorted image. This formulation moves beyond binary real/fake predictions to quantify how trustworthy a detector's decision is under distortion.

10 retrieved papers
Distortion-Aware Confidence Model (DACOM)

The authors propose DACOM, a reference-free model that fuses forensic features, no-reference image quality descriptors, and distortion-type cues to estimate detection confidence. It is trained using full-reference image quality assessment as oracle supervision but operates without references at test time.

10 retrieved papers
Two-stage pipeline for confidence estimation with severity binning and label assignment

The authors develop a training methodology that first performs type-wise adaptive binning of distorted images based on full-reference quality scores and assigns detectability labels, then trains DACOM to predict these labels using only inference-available features.

1 retrieved paper

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Distortion-Aware Confidence (DAC) formulation for forensic detectors

The authors introduce a formal definition of Distortion-Aware Confidence as a sample-level probability that a forensic detector will correctly classify a given distorted image. This formulation moves beyond binary real/fake predictions to quantify how trustworthy a detector's decision is under distortion.

Contribution

Distortion-Aware Confidence Model (DACOM)

The authors propose DACOM, a reference-free model that fuses forensic features, no-reference image quality descriptors, and distortion-type cues to estimate detection confidence. It is trained using full-reference image quality assessment as oracle supervision but operates without references at test time.

Contribution

Two-stage pipeline for confidence estimation with severity binning and label assignment

The authors develop a training methodology that first performs type-wise adaptive binning of distorted images based on full-reference quality scores and assigns detectability labels, then trains DACOM to predict these labels using only inference-available features.

Table of Contents