Multimodel_setup / docs /MULTI_MODEL_EVALUATION_PLAN.md
Thiru1122's picture
Add standardized detector adapter and evaluation documentation
6f9e017
|
Raw History Blame Contribute Delete
6.59 kB

A newer version of the Gradio SDK is available: 6.32.0

Upgrade

Multi-Model Evaluation and Comparison Plan

Target Space: https://huggingface.co/spaces/Thiru1122/Multimodel_setup
Status: Architectural Planning Document (Preparation Phase — No Agreement Layer Implemented Yet)


1. Overview & Objective

No individual deepfake detection architecture generalizes universally across all generative algorithms (GANs, autoregressive transformers, latent diffusion, flow-matching) and manipulation techniques (inpainting, face-swapping, splicing).

The objective of this framework is to establish a rigorous, standardized, multi-model evaluation harness that allows multiple independent detectors to analyze identical input assets under consistent protocols, producing standardized diagnostic outputs and agreement metrics without modifying individual model behaviors.


2. Common Input Format

To ensure parity across all candidate models, all inputs entering the evaluation harness must conform to the following pipeline:

  1. Primary Input Types:
    • PIL.Image.Image object, OR
    • Absolute file path string pointing to a local image file.
  2. Standard Ingress Transformation:
    • Validation against corruption and truncated headers using PIL.Image.open().
    • Explicit conversion to 3-channel RGB mode (.convert("RGB")).
    • Preservation of original image aspect ratio and native resolution metadata prior to model-specific preprocessing.
  3. Model-Specific Downstream Adapters:
    • Each detector adapter is responsible for applying its own exact author-defined preprocessing (e.g., specific resizing interpolation, cropping dimensions, normalization means and standard deviations).

3. Common Result Schema

All candidate detectors must implement an adapter adhering strictly to the DetectorResult schema defined in src/detector_schema.py:

@dataclass
class DetectorResult:
    detector_name: str           # Unique identifier of detector (e.g. "UniversalFakeDetect")
    model_name: str              # Backbone and checkpoint descriptor
    prediction: str              # Strictly one of: "REAL", "FAKE", "UNCERTAIN", "ERROR"
    fake_probability: float      # Normalized probability in [0.0, 1.0] that input is synthetic
    real_probability: float      # Normalized probability in [0.0, 1.0] that input is real
    raw_logit: float             # Unbounded pre-activation output logit
    confidence: float            # Confidence score in [0.0, 1.0]
    threshold: float             # Operational decision boundary (e.g., 0.50)
    device: str                  # Execution hardware ("cuda", "cpu")
    preprocessing: str           # Summary description of transform pipeline
    inference_time_ms: float     # Wall-clock execution time in milliseconds
    error: Optional[str]         # Descriptive error trace if prediction is "ERROR"
    warning: Optional[str]       # Mandatory authenticity limitation notice

Allowed prediction values are strictly bounded to:

  • REAL: Detector classified input as natural photograph.
  • FAKE: Detector classified input as synthetic / manipulated.
  • UNCERTAIN: Prediction cannot be reliably assigned (e.g., logit falls in dead-band or model disagreement).
  • ERROR: Input was invalid, corrupted, or pipeline encountered an unrecoverable exception.

4. Common Error Handling Protocol

  1. Rejection at Ingress:
    • Corrupted files, non-image formats, and truncated byte streams must produce prediction: "ERROR" with a non-null error field.
  2. No Silent Error Fallback:
    • Under no circumstances may an unhandled exception or decode failure default silently to REAL or FAKE.
  3. Hardware & OOM Isolation:
    • If an individual detector exhausts GPU memory or times out, its adapter must return an ERROR result without crashing the enclosing process or aborting sibling detector evaluations.

5. Per-Model Reporting & Aggregation

When multiple detectors evaluate a single input, the evaluation harness must report:

  1. Individual Model Records:
    • Tabular display of each detector's raw logit, calibrated fake probability, assigned binary label, and latency.
  2. Consensus & Disagreement:
    • Quantitative concordance count (e.g., 2/3 detectors predict FAKE).
    • Mean and variance of predicted fake probabilities across models.
  3. The UNCERTAIN State:
    • When constituent models produce conflicting predictions (e.g., Detector A outputs REAL with 99% confidence, while Detector B outputs FAKE with 95% confidence), the composite verdict must be explicitly designated as UNCERTAIN.
    • The system must not arbitrate conflicting predictions by simple average if the underlying feature distributions are uncalibrated.

6. Rigorous Evaluation Protocol & Benchmark Requirements

Future evaluations across multiple models must adhere to the following experimental standards:

  1. Curated Ground-Truth Benchmark:
    • Evaluation must be conducted on a balanced, labeled benchmark dataset containing verified real photographs, known GAN images (ProGAN, StyleGAN), modern diffusion images (Midjourney, Stable Diffusion, Imagen, Flux), and localized image manipulations.
  2. Separate Reporting of Error Types:
    • Metrics must report False-Real (False Negative / Missed AI) and False-Fake (False Positive / Flagged Real) error rates independently.
    • Aggregate accuracy alone is misleading on imbalanced or out-of-distribution test sets.
  3. Calibration Requirements:
    • Raw logits from distinct architectures cannot be compared directly without Platt scaling or temperature calibration on a shared validation set.

7. Fundamental Disclaimer on Model Agreement

Consensus $\neq$ Authenticity Verification:
If all detectors in a multi-model ensemble classify an image as REAL, this indicates only that none of the models detected the specific synthetic patterns for which they were trained. It does not verify that the image is an authentic photograph. Generative models developed after a detector's training cutoff may bypass all detectors simultaneously.


8. Implementation Boundary Notice

Current State: Only UniversalFakeDetect and its standardized adapter (UnivFDAdapter) are active.
Strict Boundary: The multi-model agreement layer, additional detector repositories (e.g., RAID), and ensemble fusion logic are not implemented in this phase.