Comparison between TTD (Tampered Text Detection) and CSIAD (Cross-Sample Image Anomaly Detection). The forged image is a mobile payment record.

• Left: Existing TTD relies on visual features of a single sample, which struggles with subtle tampering traces.
• Right: Although image fraud detection is limited when based on single-image visual features, reviewers can often detect anomalous images through cross-sample logical inconsistencies during manual verification processes. As illustrated in Figure  (b), the tampered region includes the account and address information because the forged payment record copies the account number from other images while altering the account name and address. CSIAD incorporates cross-sample analysis based on textual information, effectively increasing the accuracy and explainability of image fraud detection.

To better address the limitations of existing image fraud detection datasets in the financial sector, we introduce a new benchmark named CrossCred. This dataset is constructed entirely from real-world data collected in complex financial fraud scenarios, where document images are carefully forged to evade current detection technologies. CrossCred includes detailed human annotations, covering both tampered regions and corresponding explanations. Unlike previous datasets that provide only a single sample per case, CrossCred supports cross-sample inference: each fraudulent case is linked to a small group of semantically similar images, enabling the evaluation of reasoning-based detection methods. Specifically, from a collection of 400K business images, we mined 109 cross-sample anomaly cases (396 total samples, averaging 3.63 samples per case) and 109 randomly selected anomaly-free images. In total, the benchmark consists of 505 images spanning 61 distinct document types.

Examples of the Human Annotation Process. Given the LLM's pre-annotated conclusions, human annotators need to determine whether anomalies exist in the actual images and whether the LLM's pre-annotated conclusions align with the ground truth. They then select one of five predefined labels (TP, FP-A, FN, FP-N, TN) to evaluate the LLM's pre-annotated results. If the selected label is FP-A or FN (marked with *), it indicates that the LLM failed to identify the correct anomalous elements, and annotators are required to correct or supplement the anomalous elements along with the associated images. Figures (a), (b), and (c) demonstrate examples of LLM pre-annotated results as TP, FP-A, and FP-N, respectively.

Framework for CSIAD. The diagram outlines the key modules: (a) Retrieval Ensemble, (b) Sample-Specific Rule Generation, (c) Fact-Driven Verification, and (d) Rule-Based Anomaly Analysis. To determine whether a document image is fraudulent, a natural workflow is to "consult" relevant images to identify potential anomalies. As illustrated in the framework for CSIAD, which consists of four core modules:

• (a) Retrieval Ensemble: Given a query image Iq, this module retrieves a batch of relevant images, forming a suspicious set Sq.
• (b) Sample-Specific Rule Generation: Generates a set of rules Rinit that normal samples are expected to follow.
• (c) Fact-Driven Verification: Validates and filters the rules using factual evidence to yield a refined set Rverified.
• (d) Rule-Based Anomaly Detection: Uses the verified rules to detect and explain anomalies in Sq.