iPerfect

Methodology and pipeline

iPerfect separates acquisition, preprocessing, learned classification, forensic feature extraction, evidence fusion and reporting, so each stage of the reasoning chain can be inspected and tested independently.

  1. 1 · Acquisition

    Image upload or live camera capture. The file is validated and never leaves the device.

  2. 2 · Preprocessing

    Decoding, resampling to a bounded working resolution and luminance conversion for feature extraction.

  3. 3 · Feature classification

    Class-discriminative visual features drive scores for the four classes: authentic, copy-move, spliced and AI-generated/manipulated.

  4. 4 · Forensic feature engine

    Error-level analysis, EXIF/metadata verification, JPEG marker and quantisation inspection, Shannon entropy, block repetition search and a PRNU-style noise residual.

  5. 5 · Evidence fusion

    Learned and interpretable evidence are combined into a normalised score distribution over the four classes.

  6. 6 · Reporting

    The decision, confidence, every supporting indicator and a SHA-256 digest are recorded and exportable as a .ipeg container for later verification.

Forensic indicators

Error level analysisUneven recompression error across regions indicates spliced content.
Metadata verificationEXIF presence, editing-software tags and implausible provenance fields.
JPEG structureMarker sequence and quantisation tables reveal inconsistent processing.
EntropyShannon entropy of the luminance histogram flags oversmoothed or synthesised texture.
Repetition cuesDuplicated textured blocks sharing one translation vector signal copy-move cloning.
Sensor noise (PRNU proxy)Region-wise noise residual variance exposes multi-source composites.
SHA-256 hashBinds the report to the exact bytes analysed.

Interpretation and limits

Indicators are supporting evidence, not proof. Metadata can be stripped or forged, social platforms recompress uploads, and heavy post-processing weakens low-level traces.

The four classes are authentic, copy-move forgery, splicing and AI-generated/manipulated. Video, audio, legal admissibility determinations and attribution to a specific generator are out of scope.

Reported confidence describes the fused evidence for this image only; deployment robustness depends on representative data, cross-dataset testing and resistance to post-processing.