Skip to main content

Fingerprinting Technology

DigiCAP's Fingerprinting technology uses a CNN (Convolutional Neural Network)-based deep learning algorithm to extract unique features from images and videos, thereby verifying the content's authenticity.

Core Technology

The core of fingerprinting technology lies in its robustness, which allows it to track the original content without losing its uniqueness, even if the content is partially edited (e.g., cropped, resized, color-corrected) or reprocessed into another form.

Technology Overview​

Key Technologies​

Instead of relying on a single technology, this system adopts a Hybrid Model that combines two powerful approaches to maximize accuracy and efficiency.

CNN (Convolutional Neural Network) Feature Extraction

  • Quantifies the unique visual characteristics of images/videos.
  • Extracts robust feature points that are resistant to compression, resizing, and format changes.
  • Enables real-time similarity matching and tampering detection.
  • Utilizes Deep Learning-based feature extraction.

Perceptual Hash (PHash) Generation

  • Generates a unique hash value based on the overall structure and shape of image frames.
  • Effectively finds similar candidates in large-scale databases.
  • Reduces the load of CNN analysis.

Fingerprint Extraction Pipeline​

Video fingerprints are generated through the following systematic steps.

Frame-level Feature Extraction (extract frame feature)

  • First, the video is separated into individual image frames.
  • Each frame is fed into a Deep Learning-based MultiTaskModel to be converted into a 1280-dimensional CNN feature vector (fingerprint).
  • This vector compresses the core visual information of the corresponding frame.

Segment-level Feature Aggregation (extract segment feature)

  • Since frame-level features can be unstable, they are grouped into 'segments' of a certain duration (e.g., 5 seconds) and averaged (Average Pooling).
  • This creates a noise-resistant and stable representative fingerprint for the segment.

Hybrid Fingerprint Generation and Storage

  • The hybrid fingerprint is completed by combining the segment's CNN fingerprint with a 144-bit Perceptual Hash value.
  • This information is stored in a .pth file format and kept in a database.

Excellent Robustness and Accuracy​

Thanks to the abstract features learned by the AI, this system's fingerprints can accurately identify the original content despite various modifications such as:

  • Re-encoding and Compression: Robust against quality degradation and compression that occurs when uploading to various platforms.
  • Resolution and Aspect Ratio Changes: Detection is possible even if the video size is adjusted or parts of the screen are cropped.
  • Corrections and Filter Applications: Maintains the uniqueness of the original even when visual effects like brightness, contrast, and color filters are added.

For technical inquiries, please contact info.bornid@digicaps.com.