DIBench

We release simulator-generated samples from DIBench for evaluating the decision integrity of GUI-based mobile agents under deceptive injections. The samples include paired benign/injected screenshots and task annotations from food ordering (meituan), online shopping (shopping), and travel planning (travel) simulators.

Authors: Li Hu, Kanghua Mo, Yingbin Jin, Qingqing Ye, Haibo Hu.

Data format

Each line of index.jsonl contains sample_id, app, task_type, and info (the path to info.json relative to the dataset root).

Task types are Exploratory, Exact Match, Cost, Spatio-Temporal, and Quality/Popularity (stored as Quality_Popularity).

Each sample directory contains screenshots and an info.json with these fields:

Field	Description
app, task_type	Simulator and task category.
instruction	User instruction.
benign	Benign screenshot filename.
product_info	Ordered candidate records with IDs and attributes.
attacks	Paired injected variants: file (screenshot filename), attack_class (probe type), and target_index (target candidate index).
acceptable_candidate_indices	Candidate indices recorded as acceptable for the instruction.

Screenshot paths are relative to the sample directory. Candidate indices are zero-based positions in product_info, distinct from candidate IDs.

Optional evaluation.json files provide per-image anchors: conditions maps screenshot filenames to candidates, each with index, id, and bbox in screenshot pixels.

Use and metadata

The samples support decision-integrity evaluation and defense research.

See croissant.json for metadata and Responsible AI information, and LICENSE.md for licensing terms.