EVERY TAMPERED
PIXEL LEAVES A
SIGNAL TRACE.
Black-box neural networks hallucinate on re-compressed JPEGs and cannot explain why a region was flagged. PixelProof runs every image through 7 deterministic Digital Image Processing (DIP) pipelines — isolating quantization ghosts, Daubechies db4 wavelet noise cavities, lexicographically sorted DCT clones, and 2D-FFT resampling harmonics in under 1.2 seconds on CPU.
EXPLORE 6 BENCHMARK CASES OR FORGE AN IMAGE LIVE
Don't have a tampered file handy? Switch between 6 ground-truth benchmark cases (Splicing, Copy-Move Cloning, Gaussian Airbrushing, 4× Resampling, Double JPEG Phase Shift, and Authentic Camera Control) or open the Live Tamper Sandbox to clone-stamp, splice, or blur a clean photo directly on an HTML5 canvas and test detection.
UPLOAD ANY IMAGE FOR FULL 7-LAYER AUDIT
Drop any suspect JPEG, PNG, WebP, or BMP file into the dedicated analyzer. Compare the original against false-color heatmaps with a Split-Screen Slider, inspect 8×8 DCT blocks with a 5× Synchronized Pixel Loupe, tune DIP parameters ($Q$-factor, wavelet basis, morphology kernel), and export a JSON/PDF forensic dossier.
SEVEN CLASSICAL DIP DETECTORS. ZERO BLACK BOXES.
ERROR LEVEL & JPEG GHOSTS
Re-encodes the image across Q ∈ {75, 85, 90, 95} and normalizes residuals against Sobel edge energy to expose foreign spliced quantization floors.
WAVELET NOISE VARIANCE
Extracts the Daubechies db4 diagonal detail sub-band (HH1) and 5×5 Adaptive Wiener residual to flag Gaussian airbrushing and ISO mismatches.
HYBRID COPY-MOVE CLONES
Lexicographically sorts quantized 16×16 2D-DCT blocks in O(N log N) alongside ORB keypoints to isolate internal duplicated regions.
RESAMPLING SPECTRUM
Convolves luminance with an isotropic 3×3 Laplacian kernel and inspects the 2D-FFT annular ring for periodic interpolation harmonics.
DOUBLE JPEG & 8×8 GRID
Measures 8×8 Blocking Artifact Grid phase shifts (x mod 8, y mod 8) and AC(0,1)/AC(1,0) DCT histogram comb periodicity.
CHROMATIC EDGE DIVERGENCE
Computes rotational-symmetric Scharr derivatives across R, G, and B channels to detect composite boundary seams and optical aberration shifts.
DQT MATRIX & REGION MASK
Fingerprints 64-element JPEG Luminance Quantization Tables (DQT) and fuses all spatial maps via Otsu thresholding and Morphological Opening/Closing.
RUN LOCALLY OR QUERY VIA REST API
# Start the 7-Layer FastAPI DIP Forensic Engine (zero venv/env required) cd backend pip install -r requirements.txt uvicorn app.main:app --reload --port 8000 # Run the 6-case ground-truth forensic benchmark suite python scripts/evaluate_benchmark.py
6-CLASS FORENSIC BENCHMARK MATRIX
| ATTACK VECTOR | PRIMARY TRIGGER | SCORE | VERDICT |
|---|---|---|---|
| SINGLE-COMPRESSED CONTROL (Q=92) | L5: 8x8 GRID (39.6) | 27.0 / 100 | AUTHENTIC |
| FOREIGN OBJECT SPLICE (Q=74 -> Q=95) | L1: YCbCr ELA (98.0) | 83.0 / 100 | TAMPERED |
| INTERNAL CLONE STAMP (Δx=+256px) | L3: DCT+ORB (82.4) | 55.9 / 100 | TAMPERED |
| GAUSSIAN HEALING BRUSH (σ=7.0) | L2: DWT HH1 (92.7) | 71.6 / 100 | TAMPERED |
| 4x UPSAMPLED INTERPOLATED COMPOSITE | L4: 2D-FFT (96.0) | 73.4 / 100 | TAMPERED |
| PHASE-SHIFTED 8x8 CROP (5 mod 8) | L7: DQT/BAG (72.0) | 53.8 / 100 | TAMPERED |