Skip to content
tabeen.dev
Research
ActiveComputer vision

TruPhoto — Three-Class Image Authenticity Classification

Tabeen Raoof and Joao Cozza · CS483, San Francisco Bay University, 2026 · supervised by Prof. Emdad Khan

Three-class image authenticity classifier (Real / Forged / AI-Generated) combining a frozen MobileNetV2 backbone with frequency-domain features, classified by an RBF-kernel SVM.

Apr 2026

GitHub

Overview

Most published detectors treat image authenticity as binary — real versus fake. TruPhoto outputs three classes: Real, Forged (splicing and copy-move), and AI-Generated. A Photoshop manipulation and a diffusion output have different forensic signatures, and collapsing them discards information a forensic user needs. The three-class framing is a harder problem: chance is 33%, not 50%.

Approach

Two parallel feature paths feed a single classifier.

Semantic path. A frozen MobileNetV2 backbone produces 1,280-dimensional semantic features — what is in the image.

Frequency-domain path. A parallel path produces 228-dimensional content-invariant features — how the image was made:

  • FFT radial power-spectrum fitting, following Dzanic et al. (NeurIPS 2020) — 102-d
  • Block-DCT coefficient statistics, following Frank et al. (ICML 2020) — 126-d

The two paths are concatenated to 1,508 dimensions and classified by an RBF-kernel SVM. Nothing is fine-tuned; both backbones are frozen feature extractors. This is the project's actual contribution — the frequency path is what separates TruPhoto from a CNN-features-plus-classifier baseline.

Results. 75.70% test accuracy on 4,519 held-out images, a +2.43pp improvement over the CNN-only baseline of 73.27%. The frequency features alone reach 71.14% accuracy using only 228 dimensions and no semantic content at all. Training data: 30,123 images from 36 sources across ArtiFact and CASIA 2.0, spanning 25 AI generators. The full pipeline is CPU-only and runs in about 5.5 hours on a laptop with no GPU.

What the iterations showed

Accuracy decreased across the first four iterations as each shortcut was removed — roughly 84% → 76% → 77.6% → 74.9%. A cross-source sanity check surfaced three dataset-handling bugs along the way:

  1. A label-inference bug silently dropping 24 of 25 AI generator sources.
  2. Sampling that gave one generator about 70% of the AI training set.
  3. A dataset-identity shortcut — the model had learned to tell ArtiFact apart from CASIA rather than real from forged.

Iteration 5 recovered 2.43pp through the frequency path — a genuinely new representation, rather than a new shortcut.

Why we didn't fine-tune

Three independent lines of evidence pointed away from fine-tuning the backbone:

  • Per-source spread. Accuracy varied by 62.7pp for AI-Generated and 59.4pp for Real across sources, both far above the ~15pp threshold at which fine-tuning is considered safe.
  • More data didn't help. Accuracy fell when training data was increased 2.5×, indicating a representational ceiling rather than a data shortage.
  • A specific failure mode. StyleGAN3 at 37% accuracy is a representation problem, not a capacity problem — its alias-free upsampling removes exactly the spectral artifacts the DCT path is built to detect.

Limitations

  • StyleGAN3 (37%) and Stable Diffusion (62%) are both below deployment-useful thresholds.
  • JPEG storage at q=95 attenuates frequency features.
  • Generative inpainting regressed accuracy by about 10pp, because image-level frequency statistics get diluted by surrounding authentic pixels.
  • Output probabilities are uncalibrated.
  • Single-seed evaluation — no cross-validation.

Stack

Python, PyTorch, timm, MobileNetV2, scikit-learn, scipy, numpy, Pillow, Gradio

Architecture

Flowchart of the TruPhoto experiment pipeline. Images are resized and split into train, validation and test sets, then three frozen feature paths run in parallel: CNN features (Xception or MobileNetV2), frequency features (FFT radial bins and DCT statistics), and a CLIP ViT-L/14 baseline. Random-forest and SVM classifiers are trained on each path alone and on CNN plus frequency and CLIP plus frequency combinations, and all are compared in test evaluation. The CNN plus frequency path is highlighted as the final system.
Highlighted in orange: the final system — MobileNetV2 and frequency features, concatenated and classified by an SVM. Select the diagram to open it at full size.

Status

Preprint in preparation — arXiv link coming soon.

Stack

PythonPyTorchtimmMobileNetV2scikit-learnscipynumpyPillowGradio