Jevgeni Striganov
IFP — Image Fingerprint Protection | app.1ifp.com | Working prototype, proprietary algorithm
IFP introduces a structural fingerprinting approach to forensic image identification that operates without machine learning, neural networks, or AI inference. The system re-links visually modified content to registered originals through structural analysis of image composition, surviving modification combinations that defeat all major existing solutions including reverse image search engines and perceptual hashing systems. The system operates on a binary match principle: it either confirms a match to a registered original or returns no result. There is no intermediate “probable match” state. The multi-stage validation architecture produces a zero false positive rate by design — a critical requirement for forensic and legal applications. When a match is confirmed, the system reports a structural proximity score (84–99%) reflecting the degree of structural data preserved after modification, not a probability of correctness. A forensic license tracing layer tracks specific licensed copies through modification chains, enabling legal evidence generation for unauthorized usage disputes.
Digital image theft is trivial to execute and nearly impossible to prove once the image has been modified. A photographer, artist, or brand publishes an original. Someone downloads it, crops it, flips it, adjusts contrast, adds a text overlay, and uses it on their website or in advertising. The original pixels are still present — but transformed beyond the detection threshold of every existing tool.
Current solutions fail at this task:
Reverse image search systems rely on visual similarity matching. They find the original when the image is unmodified or lightly cropped. They fail on random angle rotation, hard crop, flip, B&W conversion, blur, and UI overlay. They cannot distinguish between the original and a visually similar but different image.
Perceptual hashing systems find images in crawled databases when modifications are minimal. They fail on hard crop, arbitrary rotation, perspective distortion, and combined modifications. They return 0 matches for modifications that IFP resolves at 99% structural proximity.
AI-based approaches (CLIP, learned embeddings) match semantic content, not structural identity. They answer “is this a similar image?” not “is this the same image, modified?” This distinction is critical for forensic and legal applications where provenance, not similarity, is the requirement.
No existing system provides forensic-grade structural identification of modified images combined with license tracing and legal evidence generation.
IFP is built on a fundamentally different premise: images have structural identity that persists through visual transformation.
The system uses a code-only approach — a proprietary algorithm combining established mathematical frameworks (including principles from cellular automata theory and frequency-domain analysis) with novel structural decomposition methods. No AI or ML components are integrated at any stage.
Structural, not contextual. IFP does not analyze what an image depicts. It analyzes how the image is composed at a structural level. Two photographs of the same subject taken one second apart are structurally different images. One photograph modified through crop, rotation, and color shift remains structurally the same image.
Hierarchical decomposition. The algorithm decomposes image structure into a multi-level representation. Higher levels enable fast candidate identification across the database. Lower levels provide granular structural validation, eliminating false positives. This hierarchical approach enables both speed (searching large databases) and precision (confirming identity under heavy modification).
Content isolation. Before structural analysis, the system identifies and isolates the content region within the query image. This is critical for real-world usage where the query is typically a screenshot containing browser frames, UI elements, text overlays, and other non-content artifacts. The isolation process identifies the boundary between image content and surrounding noise based on structural coherence analysis, producing a clean content region for fingerprint matching.
Multi-stage validation. Candidate matches pass through a multi-stage validation pipeline where each stage independently confirms or rejects the match. If any stage rejects, the candidate is discarded — the system returns no result rather than a low-confidence guess. This architecture produces a zero false positive rate by design. The system either confirms a match with full structural validation, or stays silent. This binary behavior is a deliberate architectural decision: in forensic and legal applications, a false match has real consequences, and a “maybe” is worse than no answer.
IFP does not perform AI content tracing. It does not detect images that were regenerated, restyled, or created from scratch using AI tools (img2img, style transfer, text-to-image). The original pixel structure must be present in the query — transformed, but present. This scope boundary is by design: IFP solves provenance (“is this my image?”), not similarity (“does this look like my image?”).
All benchmarks performed on the working prototype (app.1ifp.com) running on 2 × 16-core / 30 GB RAM servers. Query images are real-world screenshots and modified versions, not laboratory-controlled inputs.
Note on scoring: The system operates as a binary detector — it either confirms a match or returns no result. Structural Proximity reflects how much structural data survived the modification, not a probability. The false positive rate is zero by design.
| Modification | Match | Structural Proximity | Reverse Image Search | Perceptual Hashing |
|---|---|---|---|---|
| Original (unmodified) | Confirmed | 99.0% | Found | Found |
| Horizontal flip | Confirmed | 99.0% | — | — |
| Rotation (arbitrary angle) | Confirmed | 94–99% | — | — |
| Hard crop (>50% removed) | Confirmed | 84–99% | — | — |
| B&W / grayscale conversion | Confirmed | 99.0% | — | — |
| Color grading / hue shift | Confirmed | 99.0% | Partial | — |
| Brightness/contrast shift | Confirmed | 99.0% | — | — |
| Blur (gaussian, up to 45/100) | Confirmed | 84–95% | — | — |
| Perspective distortion | Confirmed | 91–96% | — | — |
| UI/text overlay | Confirmed | 92–99% | — | — |
| Noise overlay | Confirmed | 99.0% | — | — |
Table 1. Single modification detection results. “—” indicates the system returned no match.
| Combined modifications | Match | Structural Proximity |
|---|---|---|
| Crop + flip + contrast | Confirmed | 99.0% |
| Crop + flip + contrast + noise + UI overlay | Confirmed | 99.0% |
| Hard crop + hard UI overlay | Confirmed | 99.0% |
| Hard crop + hard UI + inside another UI (recursive) | Confirmed | 85.0% |
| Noise + contrast + arbitrary angle + overlay | Confirmed | 99.0% |
| Greyscale + noise + contrast + overlay | Confirmed | 99.0% |
| Screenshot of search results containing the image | Confirmed | 99.0% |
Table 2. Combined modification detection results.
The system correctly distinguishes between structurally identical (modified original) and visually similar but different images. Two images of similar visual appearance but different structural origin are not matched. This is verified through the high-similarity detection test cases documented in the demo.
Beyond origin identification, IFP embeds forensic license identifiers that survive the same modification spectrum as the structural fingerprint. Each licensed copy receives a unique identifier (e.g., LIC-QVVE13, LIC-5I587H) that persists through:
This enables tracking of specific licensed copies through modification chains — identifying not just that an image was copied, but which specific licensed copy was the source of the unauthorized use.
Concept — In Development
IFP includes a concept for automated ownership dispute resolution. When a registrant claims ownership of an image already registered by another user, the system compares structural evidence (image size, bit depth, format quality, color profile, creation date, camera EXIF, XMP history, metadata richness, embedded thumbnail presence, and resampling detection) to determine which image is more likely to be the original.
Known limitation: Metadata-based dispute scoring is inherently unreliable in certain scenarios. For example, a PNG screenshot of a JPEG original will contain higher bit depth and data quality than the original itself — producing a false signal that the copy is the original. This fundamental ambiguity in metadata comparison makes the current dispute mechanism insufficient for forensic-grade decisions. The dispute system is considered experimental, and C2PA provenance verification is being explored as a more reliable foundation.
When a C2PA manifest is attached to an image during original registration, IFP stores it as part of the registration record. If a copy is later submitted for registration, the system cross-references the C2PA manifest of the already-registered original to reject the duplicate before it enters the database. This prevents ownership hijacking at the registration stage. C2PA manifest integration into the full dispute resolution pipeline is the primary direction for making ownership disputes forensically reliable.
Creative IP protection. Photographers, illustrators, and brands register originals before publication. Unauthorized modified usage is detected through screenshot upload and matched to the registered original with forensic-grade confidence.
Legal evidence generation. Match results include structural proximity score, registration timestamp, license data, and modification survival documentation — producing evidence suitable for DMCA takedowns, cease-and-desist, and litigation.
Content licensing enforcement. Publishers and stock platforms can trace specific licensed copies to detect license violations even when the licensed image has been heavily modified.
Platform integrity. Social media and advertising platforms can verify whether uploaded content matches registered originals in the IFP database, enabling automated detection of unauthorized content use at scale.
IFP is a working prototype with a functional web application demonstrating full registration, detection, license tracing, and dispute capabilities. The algorithm has been independently developed by a single researcher over the course of applied R&D.
Areas of active development:
This document describes system capabilities and benchmark results. The underlying algorithm, implementation details, and processing pipeline are proprietary. Technical details are available under NDA for qualified research institutions and commercial partners.
