Aletheia: Advancing Image Forensics for Adversarial Data Detection

Executive summary

Aletheia helps users detect manipulated imagery and analyze the likely nature of adversarial changes. It combines ensemble model techniques, heatmap overlays, adjustable sensitivity, and reporting workflows to improve confidence in image evidence.

Challenge

AI-enabled image manipulation has made deepfakes, spliced imagery, and poisoned datasets easier to produce. These threats can compromise intelligence, investigations, public trust, and national-security decision-making.

• Existing forensic tools often focus on narrow manipulation types.
• Manual analysis can be slow and difficult to scale.
• Decision-makers need explainable outputs, not only a binary detection result.

Problem requirements

An effective platform needed to detect multiple manipulation types, scale across large datasets, integrate into analyst workflows, and provide transparent outputs that support confident decisions.

Solution

Aletheia uses an ensemble of convolutional neural networks and a metaclassifier to identify copy-move manipulation, splicing, and inpainting. The interface supports batch uploads, progress visibility, heatmaps, manipulation summaries, and downloadable reports.

• Copy-move detection for duplicated and relocated image regions.
• Splicing detection for inserted elements from other images.
• Inpainting detection for removed or fabricated image areas.
• Heatmap overlays and adjustable sensitivity for explainable review.

Implementation

VISIMO trained Aletheia on more than 120,000 images representing varied manipulation methods. The application uses role-based access and dashboard views so users can inspect probabilities, heatmaps, and detailed analysis summaries.

Results

Aletheia improved detection coverage across tested manipulation types, supported batch workflows, and reduced analysis time while broadening the platform's applicability beyond defense use cases.