# DeepImageSpam: Deep Learning based Image Spam Detection

**Authors:** Amara Dinesh Kumar, Vinayakumar R, Soman KP  
**arXiv:** [1810.03977](https://arxiv.org/abs/1810.03977)  
**Date:** 2018-10-03  
**Summary page:** https://www.summarizepaper.com/en/paper/1810.03977v1/

## Key points

- Growing concern of hackers and spammers using innovative techniques to deceive internet users
- Image spam manipulation makes it indistinguishable from the original image
- Deep learning approach using convolutional neural networks (CNNs) proposed for image spam detection
- Dataset used consists of 810 natural images and 928 spam images for classification
- CNN model trained with binary entropy loss and adam optimizer, saving checkpoints in hdf file
- Training and testing performed on distributed computing cluster platform with i7 CPU processor and 8 GB RAM system configuration
- Libraries used include Keras, Sklearn, and Tensorflow deep learning
- Proposed deep learning approach achieves accuracy of 91.7%, outperforming existing techniques
- Metrics evaluated include accuracy, recall, precision, and F1 score
- CNNs provide better accuracy compared to traditional machine learning methods for image spam detection
- Deep learning approach eliminates manual feature extraction by automatically identifying features, reducing time and effort
- Suggestion for further research exploring other deep learning algorithms like RNN and LSTM, as well as tuning architecture and hyperparameters to improve image-related techniques

## Summary

The paper titled "DeepImageSpam: Deep Learning based Image Spam Detection" addresses the growing concern of hackers and spammers using innovative techniques to deceive internet users. One such technique is image spam, where spammers manipulate portions of an image to make it indistinguishable from the original, fooling users. To combat this issue, the authors propose a deep learning approach using convolutional neural networks (CNNs) for image spam detection. The authors utilize a dataset consisting of 810 natural images and 928 spam images for classification. They train the CNN model with binary entropy loss and adam optimizer, periodically saving checkpoints in an hdf file. The training and testing are performed on a distributed computing cluster platform with an i7 CPU processor and 8 GB RAM system configuration. The libraries used for training and testing include Keras, Sklearn, and Tensorflow deep learning. The results show that the proposed deep learning approach achieves an accuracy of 91.7%, outperforming existing image processing and machine learning techniques. The metrics evaluated on the test dataset include accuracy, recall, precision, and F1 score. In conclusion, the study demonstrates that CNNs provide better accuracy compared to traditional machine learning methods for image spam detection. The deep learning approach eliminates the need for manual feature extraction by automatically identifying features, reducing time and effort. The authors suggest further research exploring other deep learning algorithms like RNN and LSTM as well as tuning architecture and hyperparameters to gain more insights into improving image-related techniques. Overall, this paper contributes to addressing the challenge of detecting image spam by proposing an effective deep learning-based solution that surpasses existing approaches in terms of accuracy.