Using Deep Learning to Spot False Product Photos in Online Sales

Ways to shop for various items without stepping outside became totally feasible with the e-commerce site boost mod, and it helped change the dynamic within the retail industry. However, this online facility has also led to the penetration of counterfeit and fake goods in the market, which have posed a significant threat to consumers and manufacturers. Counterfeit product images detection on e-commerce web sites is a challenging task as counterfeit products are visually highly similar to the genuine products. Due to the large volume of products in the market and a wide spectrum of means of differentiating authentic from inauthentic products, traditional means to identify them (mainly requiring human inspection or rule-based systems) are also changing, especially for counterfeit products where these means are rapidly losing effectiveness. A lot of deep learning image classification techniques is lately proposed as a good starting point for automating the detection of fake product images. In this paper, we propose to use deep learning models to address detection of counterfeit images of products on e-commerce websites. ABSTRACT. Specifically, we look at Convolutional Neural Networks (CNNs) for image feature extraction and Generative Adversarial Networks (GANs) for image verification. A CNN which has been widely studied for its power to learn spatial hierarchies in images would work in detecting the visual discrepancies and irregularities between images of genuine and counterfeited products. On the other hand, GAN s can generate fake product images for the defined model to learn the small details differentiating the real items from the counterfeits, through comparing the input real-world images with adversarially generated ones. An important element of the solution implementation is the creation of a bulk dataset, which includes product photos from different categories, specifically electronics, clothing, and accessories. The models for this approach are trained on a set of thousands of real/counterfeit images. Data Preprocessing Data preprocessing is a series of techniques, such as image normalization, augmentation, and resizing for increasing the diversity of the dataset and minimize overfitting. Also, transfershiping techniques are employed using state of the art pre-trained models such as VGG16 and ResNet which can be fine-tuned for fake product recognition hence reduces the requirement of huge training dataset We assess our deep learning models on various performance measures like Accuracy, Precision, Recall, and F1 score. And an in depth comparison of the CNN based models, GAN based models, and Hybrid models by utilizing benefits of both architecture.