Phizon

Inspiration

The base of any illicit activity on the Internet is a malicious website. The attackers can gain partial or complete control of the machine using these linkages. As a result, target systems become readily infected, allowing attackers to use them for a variety of cyber-crimes such as credential theft, spamming, phishing, denial-of-service assaults, and other similar attacks. The rising issue related to spamming, phishing and malware, has created a requirement for a solid framework solution that can analyze the extracted features, classify and further recognize the malicious URL as well as images of clone websites.

What it does

URL Check

• Adversarially attack the data using FIGA (Feature Importance Guided Attack) and create 5 different attack sets of data based on certain parameters
• We utilize these data features to implement adversarial training as a defence against FIGA using neural net architecture in PyTorch
• Web application demonstrates URL prediction in 2 forms where it serves as a standalone URL checker and also fetches random topic-based tweets to identify any malicious redirection.

Clone/Spoof Check

Upload a website screenshot/image to see whether it is a spoof of any authorized popular website to prevent sensitive information retrieval and data breaches.

How we built it

What is FIGA

FIGA is model agnostic, it assumes no prior knowledge of the defending model's learning algorithm, but does assume knowledge of the feature representation. FIGA leverages feature importance rankings; it perturbs the most important features of the input in the direction of the target class we wish to mimic.

Creating an adversarially hardened model

We train the defending model using unmodified data and correctly labelled adversarial samples. The expectation is that training the model with adversarial samples will improve its performance.

Additional capability

Added clone classification feature to classify spoof websites using deep learning model. Also, review tweets in real-time to see if there is any malicious content being posted on the social networking site.

Challenges we ran into

• Training a deep learning model on new Amazon EC2 DL1 instances using Gaudi Accelerators.
• Deciding on the parameters for data perturbation
• Improving model for decent predictions

Accomplishments that we're proud of

• Creating something that can be utilized right away for cybersecurity purposes.

What we learned

• Reading through various research papers and increasing our knowledge in the field.
• Learned more about Gaudi Accelerators, its efficiency and capabilities.

What's next for Phizon

• We aim to make a package out of it so that the model can be utilized by social media and similar applications to moderate content such as promotions and offers which may lead to unethical data extraction by fooling people with phishing urls.
• Further, improve on the model to generate better results and predictions

Built With

• amazon-ec2
• amazon-web-services
• dl1
• gaudi
• geoip2
• habana
• heroku
• python
• pytorch
• sqlite
• streamlit