Real and fake IDs: diverse ID samples (real and fake) through partnerships with government agencies, DVSA, and publicly available datasets, including Government ID verification data from countries around the world (see Persona). These should cover different types of IDs (passports, driver’s licenses, temporary licences, ID cards, and student IDs) and from various countries (international IDs).

Facial data from publicly available biometric datasets (see CelebA dataset) 

Other data sources: UK demographic data from a partnership with banks, including date of birth and names of individuals under 35, for cross-referencing with text captured in ID images.

Feature extraction from ID image (card reader input): When an ID is presented for verification, the ID image is processed by the card reader using a CNN that extracts relevant security features (like the hologram, country flag, user signature, and background visual pattern) and other design elements like image dimensions

Text extraction and parsing: Using optical character recognition (CNN), text (e.g., name, date of birth, and ID number) is extracted from the ID image and then cross-referenced with available databases for verification.

Biometric facial features extraction from ID photo (car reader input) and real-time picture (camera capture): If the given ID type contains an ID photo of the individual, the camera also captures a real-time picture of the individual. That picture is normalised and preprocessed to match the conditions and quality of the ID photo (e.g., image dimensions, resolution, brightness, contrast, and angle) to ensure consistent input data and more accurate feature extraction and model processing. 

Facial features are extracted from the ID photo and real-time image in parallel networks that leverage a pre-trained biometric facial recognition model (e.g., a CNN, such as FaceNet or DeepFace). Both CNNs share (normalised/comparable) weights to generate face embeddings, which are then used to calculate facial likeness/similarity scores and determine if faces match. 

ID image augmentation: If needed, ID images (real and fake) can be duplicated and augmented by adding visual noise and texture to simulate various real-world scenarios and increase dataset diversity (images under different lighting and weather conditions, old cards, scratched, dirty, discoloured, etc.).  This can help achieve balanced datasets of fake-to-real samples, ID types, and other key parameters.

We could engage a community of superminds (expert fake ID makers) to submit anonymous fake IDs (fake names and blurred profile pics).

Data labelling: real and fake IDs can be labelled with country, ID type, and other relevant features.

For training the GAN, the processed data will need to be split into small datasets by ID type using an initial ID-type classification model (the user can also manually indicate this in the early stages to teach the model what input to expect for each ID type).

Per ID type classifier, datasets are split into equal batches of training and testing and split again into equal batches to be fed into the Generator and Discriminator networks, respectively.

For each ID type, there is a dedicated GAN model with multiple discriminator networks per GAN, each working on several areas of the ID features (as seen in the example of Wav2Lip). Each discriminator can criticise different aspects of the documents (e.g., format, visual texture, and image) and can be trained and enhanced incrementally and separately/in parallel with other critics.

Generator: Train the generator to produce synthetic IDs from existing samples that span different real-world conditions, which can help improve network accuracy and robustness.

Discriminator: For each input received from the testing batch or the synthetic IDs, real vs. fake IDs are classified by recognising unique features, text accuracy, and facial matches.

All features are passed through the discriminator network as input to determine whether the ID is authentic.

The main discriminator outputs the final result (real or fake) based on the integrated features, including ID-specific attributes, face matching score, and text consistency.

The network outputs a verdict: real or fake ID, and “underage” or “legal age” using a local UK age threshold, such as 18 (for alcohol and clubs) or 21 (for specific alcohol sales).

Optimising the model: If the output is wrong, we could experiment with tuning hyperparameters (e.g., learning rate or batch size), or adjust the images (e.g., if underage and legal-age sets are unequal or certain weather conditions are not accounted for). Then, the training can be repeated with this new and improved batch until the output error rate is ethically tolerable in the given context/business (e.g., high-security clearance, high-risk environments such as banking systems). 

For the specific use cases of age-restricted product sales and venue access, the rate might be tolerable at 5% (this is an arbitrary number, but generally it’s a low-risk scenario and would still significantly reduce underage access). Human oversight can be used to provide their best judgment and feed back this information to the system in real-time.