BEYOND DETECTION: UNCERTAINTY, QUALITY, AND MULTI-SOURCE REASONING FOR OPERATIONAL MULTIMEDIA FORENSICS

Synthetic Image Detection and Localization Challenge is an evaluation effort focused on advancing methods for identifying and localizing synthetic and manipulated images under realistic, real-world conditions. The challenge is broadly informed by reliability-aware media forensics principles, emphasizing not just whether a system can detect manipulation, but how robustly and transparently it performs when faced with uncertainty. Participants will be evaluated on their ability to handle diverse and evolving image manipulations, generalize beyond known data and tools, and provide meaningful confidence or localization signals where possible. Building on prior community challenges at IH&MMSec, ICCV, and WACV, this effort aims to expose the gap between controlled benchmark performance and operational readiness, while fostering a research community focused on trustworthy and deployable image forensics. Top performers may be eligible for follow on research grants sponsored by the ULRI Digital Safety Research Institute.

Submit for publication in IEEE Xplore ICIP 2026 Workshop Proceedings. We will run a rigorous peer-review process with at least two expert reviews per paper and transparent conflict-of-interest handling, aligned with ICIP standards, and we aim for inclusion in the ICIP 2026 Workshop Proceedings on IEEE Xplore.

Morning

• Keynote (30 min)
• Invited Talk 1 (30 min)
• Technical Talk Session 1 (60 min):
  • Robustness Under Real-World Degradation & Dynamic Weighting of Localized Evidence (3 talks)
• Technical Talk Session 2 (60 min):
  • Uncertainty Quantification & Conformal Prediction for Forensic Reasoning (3 talks)
• Poster & Software Demo “preview” + networking (30 min)

Afternoon

• Invited Talk 2 (30 min)
• Competition block (ULRI Digital Safety Research Institute challenge)
  • Challenge overview, motivation, protocol (sequestered evaluation, unseen generators/manipulations), preliminary findings (30 min)
  • High-performing teams share results and technical approaches: lightning talks / short-paper spotlights (60–75 min)
  • ULRI DSRI highlights challenge insights and top performer recognition; next steps in independent evaluation (15 min)
• Closing discussion / practitioner feedback loop (15–20 min)
  

Quality-to-Confidence Modeling
(with dynamic evidence weighting and dataset shift)
Methods that map forensic evidence quality (e.g., compression, blur, occlusion, missing metadata, provenance gaps) into decision confidence for tasks like association, attribution, and event reconstruction. Emphasis on dynamic weighting and explicit handling of dataset shift across devices, codecs, environments, and acquisition pipelines.

Calibrated Uncertainty
(utilizing conformal prediction and proper scoring rules)
Approaches that produce well-calibrated probabilities or prediction sets suitable for forensic reporting and thresholded decision-making. Includes conformal prediction, calibration under shift, and evaluation via proper scoring rules and calibration diagnostics relevant to false-link / false-ID risk.

Multi-Source Evidence Fusion
(using evidence graphs)
Fusion techniques for combining heterogeneous sources (imagery, video, metadata, sensor/context signals, OSINT) while representing provenance, dependencies, and contradictions using graph-based evidence models. Focus on traceability, conflict resolution, and explainable aggregation of support for competing hypotheses.

Operational Robustness
Robust pipelines that remain reliable under real forensic constraints: low quality, partial observability, missing modalities, adversarial or manipulated content, and resource/latency limits. Includes monitoring, fallback strategies, and robustness to real-world acquisition variability, tampering and generators unseen in training sets.

Competition: Reliability-Centric Open Evaluation Challenge
The competition evaluates image forensic systems under realistic, operational conditions. Participants must detect and localize manipulated content. Evaluation uses sequestered data and includes previously unseen generators and manipulation types to mirror real media-authentication scenarios. Submissions are expected to remain robust to post-processing and train–test mismatch, and to output calibrated confidence scores with interpretable evidence, including an estimated manipulation type.