Source-linked AI summary
Multi-Tool Image Editing Attribution in Facial Forgery
Sheng Liu, Qiang Sheng, Danding Wang, Yu Li, Chenming Zhou, Juan Cao
TL;DR
Existing attribution methods assume a single editing tool, but facial images are increasingly edited through multi-tool sequences whose traces overlap or overwrite one another. The paper introduces MIEA, builds the MultiEdit dataset, and proposes DPEC to recover multiple tool traces; DPEC outperforms the evaluated baselines, including on five-step edits.
Problem
Multi-tool facial editing creates composite and overwritten traces that existing single-tool attribution methods cannot fully identify, motivating attribution of all involved tools.
Method
The paper formulates MIEA, constructs MultiEdit with over 500k edited facial images covering six tool types, and develops DPEC using spatial-frequency patchwise enhancement with error-based curriculum learning.
Results
DPEC outperforms all baselines across difficulty levels, achieving 92.93% mean strict accuracy, 98.56% mean toolwise accuracy, and 87.40% strict accuracy for five-step edits.
Takeaways & Limitations
MIEA and MultiEdit support more complete facial image forensics by targeting the multiple tools involved in an image's editing chain.
Takeaways & Limitations
The paper identifies inter-tool commutativity as an unresolved issue and notes that predicting edit order is future work.
Abstract
from arXiv · showhide
As generative AI tools become increasingly powerful and easy to use, people can easily edit portrait images with a prompt, necessitating the task of image editing attribution, which predicts the involved editing tools from the given image. Existing attribution methods hold the single-tool assumption and can only attribute a specific editing tool, but struggle to handle the more complex and increasingly common multi-tool editing scenarios, where artifacts left by different editing tools are composite and overlapped. To address this gap, we explore Multi-Tool Image Editing Attribution (MIEA), which aims to identify multiple editing tools involved in a multi-tool edited facial image. To simulate the real-life editing operations on facial images, we then construct a new dataset, MultiEdit, which contains 500k+ edited facial images and covers six types of editing tools that support face swapping (Deepfake) and various facial enhancements. Inspired by the findings from data analysis, we design DPEC, a multi-tool attribution method that can capture distinguishable, locality-aware editing tool traces from both spatial and frequency domains with the support of an error-based curriculum learning strategy. Experiments show \Method\ outperforms nine methods for facial images edited in at most five steps.
1. Introduction
AI image editing is increasingly accessible, while multi-tool facial editing creates overlapping traces that existing single-tool attribution methods cannot fully identify. The paper formulates MIEA, constructs MultiEdit, and proposes DPEC to recover the tools involved in an editing chain.
- Motivation: 71% of social media images are now AI-generated or AI-edited, increasing the accessibility and prevalence of synthetic image manipulation.Generative AI tools allow ordinary users to create or manipulate vivid images with prompts.
- Challenge: Multi-tool editing combines artifacts from different tools, violating the single-tool trace assumption underlying existing attribution methods.Sequential edits can mix, overwrite, or degrade earlier tool traces, making single-tool predictions incomplete or ineffective.
- Task: Multi-Tool Image Editing Attribution identifies multiple editing tools involved in a multi-tool-edited facial image rather than only one tool.The task targets the full editing process from the final edited result image.
- Dataset: MultiEdit contains over 500k edited facial images covering six representative types of editing tools.The dataset is designed to support study of multi-tool facial editing scenarios.
- Method: DPEC extracts locality-aware tool traces from spatial and frequency domains using patchwise enhancement and error-based curriculum learning.Its curriculum progresses from easier images with fewer editing steps to harder multi-step cases.
2. Related Works
Existing facial attribution methods generally assume a single model or tool fingerprints an image, whereas multi-tool editing can mix and overwrite traces. Related work has begun examining mixed traces, but the multi-tool attribution problem remains distinct.
- Single-tool attribution: Contrastive attribution methods such as RepMix and DNA-Det usually assume that an image is fingerprinted by only one model.This assumption does not match sequential editing with multiple tools.
- Trace interference: Multi-tool editing may mask or overwrite earlier traces, rendering single-tool attribution methods ineffective for recovering the complete editing history.For example, a later filter adjustment can destroy traces left by an earlier editing model.
- Prior multi-tool work: Fakechain studied mixed-trace effects on detection methods, while MSA targets involved tools between two sequential editing stages.These efforts indicate growing attention to sequential or mixed editing scenarios.
3. Dataset Construction
MultiEdit simulates sequential facial editing with six tools spanning face swapping and beautification, multiple editing targets, and varied tool combinations. Starting from 200 real portraits, the construction enumerates valid tool-use plans and produces 510k edited images.
- Selected Editing Tools: Six tools based on five model or algorithm frameworks were selected: OpenCV, FaceFusion, DiffSwap, CSD-MT, SHMT, and FLUX.1-Kontext-[dev].FaceFusion and DiffSwap support face swapping, while OpenCV, CSD-MT, SHMT, and FLUX support facial beautification operations.
- Editing Types and Targets: The dataset covers face shape, skin, and makeup editing types with targets representing frequent daily editing operations.Editing types were selected according to the practical effects and intended use scenarios of the tools.
- Editing Plans: MultiEdit varies editing types, tools, targets, and sequential tool positions to cover realistic combinations of editing operations.These combinations are intended to simulate diverse real-world multi-tool editing scenarios.
- Sequential Generation: Deepfake tools are restricted to the first step, and each generated sequence uses k = 4 beautification steps with random sampling without replacement.The ordering constraint follows the described practical application logic of performing face swapping before facial enhancement.
- Source Images: 200 real portrait images were used as source images, and unsuccessful editing cases were filtered after enumerating valid tool-use sequences.Each source image was processed under the valid plans specified by the generation algorithm.
- Dataset Overview: 510k edited images were produced, with each source image processed through 648 valid plans.The resulting dataset statistics are summarized in Table 3.
13 end
The analysis examines detectability, regional consistency, and domain diversity in multi-tool editing. It finds that spatial and frequency traces provide complementary evidence, while perceptual change does not necessarily increase steadily with editing steps.
- The analysis evaluates single-tool detectability, domain diversity, regional effects, and inter-tool interactions in MultiEdit.
- Multi-tool editing is evaluated by training binary classifiers on single-step edits and testing whether they detect specific tools after subsequent edits.
- CSD-MT, SHMT, and DeepFake overlap spatially but differ in frequency, whereas OpenCV and FLUX show the opposite pattern.
- Spatial and frequency domains offer mutually complementary information for multi-tool image editing attribution.
- Regional editing effects are analyzed through low-, mid-, and high-frequency DCT bands across facial regions, with significance tested using the Kruskal-Wallis H test.
- LPIPS scores from steps 2 to 5 do not increase continuously, contrary to the expectation that more editing steps steadily increase perceptual loss.
4. Proposed Method: DPEC
DPEC formulates multi-tool image editing attribution as multi-label prediction and combines dual-domain, patchwise feature enhancement with error-based curriculum learning. Its curriculum progressively shifts training toward harder samples using error feedback.
- Problem Formulation: Multi-tool attribution predicts, for each candidate tool, its probability of having been used in editing an image.The model outputs an n-dimensional vector whose entries correspond to tools in the candidate set.
- Method Overview: DPEC combines Patchwise Feature Enhancement and Error-based Curriculum Learning for fine-grained facial-image attribution.PFE captures locality-aware traces, while ECL controls the difficulty of training samples.
- Patchwise Feature Enhancement: Patchwise Feature Enhancement uses sliding-window DCT, spatial-frequency cross-attention, and position awareness to extract tool-specific local features.It selects distinctive patches and enhances intra- and inter-patch features across spatial and frequency domains.
- Error-Based Curriculum Learning: Error-Based Curriculum Learning partitions training images by editing-step count and dynamically adjusts subset sampling weights from validation error rates.Subset S_i contains images edited with i steps, and the sampling weights are updated using a self-paced learning objective.
- Error-Based Curriculum Learning: As training accuracy increases, ECL increasingly incorporates hard samples with more editing steps and non-linearly mixed effects.The strategy begins with easier samples containing more clean traces and progresses toward harder samples based on error feedback.
5. Experiments
Experiments evaluate DPEC for multi-tool facial-image attribution across difficulty levels, robustness settings, component variants, editing-area patterns, and region attention. DPEC generally outperforms the selected baselines, while performance declines as editing traces increasingly overlap.
- Main Results: DPEC achieves 92.93% mean strict accuracy and 98.56% mean toolwise accuracy, surpassing all baselines and state-of-the-art methods.On images with five editing steps, DPEC reaches 87.40% strict accuracy.
- Main Results: As editing steps increase, all models degrade because traces from different tools overlap and mask one another.Single-tool methods are additionally limited because they attribute observed features to only one tool.
- Robustness: DPEC retains a significant performance advantage over baselines after compression and resizing, despite lower accuracy than on uncompressed images.These transformations model common social-media processing conditions.
- Ablation Study: Removing either PFE or ECL causes a significant performance drop, supporting the importance of both feature extraction and progressive learning.The ablation evaluates variants without PFE, without ECL, or without both modules.
- Ablation Study: SWDCT outperforms standard DCT and DWT variants by approximately 14.7% and 7.1%, respectively, across editing-step settings.The result supports SWDCT as a more effective source of distinguishable frequency-domain features for MIEA.
- Further Analysis: Large-area edits in later steps make earlier tools harder to identify, yet DPEC reaches 81.32% Acc-S on small-to-large samples.Large later edits can overwrite or obscure traces left by previous tools; no small-to-small samples were available.
- Further Analysis: Grad-CAM shows tool-specific facial regions, while some tools also rely on adjacent non-facial regions when facial traces overlap.Figure 8 visualizes original images, multi-tool edits, and tool-specific heatmaps with editing targets marked.
6. Conclusion
The paper introduces MIEA, MultiEdit, and DPEC to identify multiple editing tools in facial images. It reports DPEC’s superiority in multi-tool attribution and identifies sequence prediction and general-domain evaluation as future directions.
- Conclusion: MIEA identifies multiple editing tools involved in a single edited image, making the editing process more transparent and supporting accountability for image authenticity.The task extends attribution beyond identifying only one editing tool.
- Conclusion: MultiEdit is a 500k-scale dataset covering combinations of six editing tools, and DPEC learns individual tool features across two domains.DPEC also uses error-based curriculum learning to progress from easy to hard samples.
- Conclusion: Extensive experiments show DPEC’s superiority in multi-tool attribution, while future work will explore tool-use sequence prediction and general natural images.The current task focuses on identifying the involved tools rather than their order.
Supplementary Material
The supplementary material contains limitations, implementation details, edited-image cases, and additional analyses. It covers Qwen3-VL evaluation, DPEC design and training, sensitivity, failure cases, and sequence recovery.
- Supplementary Organization: The supplementary material begins by stating that its contents are organized into subsequent sections.This introductory statement provides the supplement’s organizational transition.
- Supplementary Organization: Section A covers limitations and future work, while Section B provides implementation details for Qwen3-VL evaluation, DPEC design, and training.The material also includes edited-image cases and additional analyses.
- Supplementary Organization: Additional analyses address key hyperparameter sensitivity, failure cases, and attempts at sequence recovery.These analyses are listed under Section D.
A. Discussion on Limitations and Future Work
The discussion describes the paper’s infrastructure for multi-tool facial-image attribution and identifies scope boundaries in dataset coverage, tool-order prediction, and closed-set attribution. It also documents implementation choices and Qwen3-VL validation.
- Discussion: The work formulates MIEA, builds a large-scale multi-tool-edited dataset, implements nine baselines, and introduces a new method.It frames the contribution as an extension of deepfake detection and attribution for multi-tool editing.
- Limitations and Future Work: MultiEdit covers human-face editing but not general-purpose image editing, limiting direct coverage of non-facial image domains.The authors plan to extend the dataset to general edited images.
- Limitations and Future Work: The task predicts the set of involved tools rather than their use order; order prediction was deferred after initial attempts produced no reasonable results.The authors speculate that some tool effects may be partially commutative and leave further study for future work.
- Limitations and Future Work: The closed-set formulation restricts attribution to preset tool candidates, so predictions for images edited by unseen tools are unreliable.The authors plan to investigate open-set attribution for more complex real-world scenarios.
- Implementation Details: Manual sampling found over 90% consistency between Qwen3-VL’s editing-success judgments and human evaluations.Qwen3-VL was used to determine whether editing operations had been successfully applied.
- Implementation Details: DPEC’s implementation uses SWDCT for spatial-to-frequency transformation with a 16×16 window and step size 2, while ECL uses polynomial order 3.The cross-attention module uses 8 heads and hidden size d_k = 768.
B.3. Details of Training Process
The training process uses distributed optimization with warmup followed by cosine annealing, while Qwen3-VL-CLS uses LoRA and binary cross-entropy for multi-label classification.
- Training uses Accelerate for distributed training and initializes the learning rate at 3e-4.
- The learning rate increases linearly during warmup according to the current epoch and warmup epochs.T and T0 denote the current and warmup epochs, respectively.
- After warmup, cosine annealing decays the learning rate using total training epochs and a multiplicative factor of W × 0.99 + 0.01.The cosine decay factor is based on progress from the end of warmup to the total remaining training period.
- Qwen3-VL and Qwen3-VL-CLS use Qwen3-VL-30B as the base model.For Qwen3-VL-CLS, LoRA trains both text and image encoders.
- Qwen3-VL-CLS uses binary cross-entropy loss for multi-label classification.
- MultiEdit samples include multi-tool edits on real faces, with additional edited samples shown in the supplementary figures.The dataset retains some samples with relatively heavy makeup to cover more diverse scenarios.
D.1. Results on Hyperparameter Sensitivity
Ablation studies evaluate PFE and ECL operations and hyperparameters, finding that DPEC remains highly effective across non-optimal settings, especially beyond two-tool edits.
- Ablations examine the key operations and hyperparameters of PFE and ECL.The results are reported in Table S1.
- DPEC shows a degree of hyperparameter insensitivity, maintaining high performance on images edited by more than two tools despite non-optimal settings.
D.2. Failure Case Analysis
DPEC can miss earlier tools when later large-area edits obscure traces from preceding small-area edits. A preliminary sequence-level formulation achieves limited exact-match accuracy, indicating that recovering editing order remains difficult.
- Failure Case Analysis: DPEC fails to detect some tools in sequentially edited images when preceding traces are obfuscated by later edits.
- Failure Case Analysis: Large-area later edits can cause missed detection of preceding tools, and pixel-level differences between consecutive edits influence trace detectability.
- Tool-Use Sequence Recovery: The main task performs set-level multi-label attribution without modeling tool-use order.A preliminary extension instead attempts sequence-level tool-use recovery.
- Tool-Use Sequence Recovery: The sequence-level representation uses an N × T matrix in which M[i, j] = 1 indicates Toolj at editing step i.N denotes the maximum editing steps and T the total number of editing tools.
- Tool-Use Sequence Recovery: 12.64% average EM is achieved across test samples, while EM reaches 1.67% for images edited by 5 tools.Exact match requires both the tool set and its order to be correct.