get real: real vs fake image detection

generative ai can produce realistic product images in seconds, which lowers the cost of building scam websites and running social engineering attacks. the goal was to build a model that classifies e-commerce product images as real or ai-generated, so it could plug into a broader scam-detection pipeline alongside signals like domain registration and ip activity.

i built a balanced 6,000-image dataset from scratch. 3,000 real product images were sampled from the amazon berkeley objects dataset, then run through google gemini pro to generate a one-line caption of each object.

those captions were fed to dall-e 2 to generate 3,000 'like-for-like' fake images - the same products, but synthetic. this gave clean, balanced classes and a variety of product types against varying backgrounds, so the model would generalise across the range of images a genai model can produce.

the synthetic images are convincing enough to pass as real product photos on a storefront, which is the core of the detection problem.

i used transfer learning on three families of pretrained models - vgg-19, resnet50, and efficientnet (b0, b3, b7) - freezing the backbones and training only the output layer. everything was trained on an a100 gpu on google colab over 20 epochs.

efficientnetb7 reached 99.44% accuracy, with resnet50 and the efficientnet variants close behind. vgg-19 lagged, with a noticeably higher false-negative rate.

to mimic real-world conditions, i tested the models on unseen dall-e 3 images. accuracy dropped sharply: vgg-19 fell to 24%, resnet50 to 48%, and efficientnetb0 was best at just 50%. efficientnet's squeeze-and-excitation blocks helped it adapt, but the drop is a clear sign of how hard it is to keep detectors current as genai improves.

finally, i embedded the models into a chrome extension so you could check whether images online are real or fake in the browser.