ArmourAI | Devpost

Armour AI

Inspiration

Our journey began at the University of Cambridge GenAI Hackathon, inspired by a talk from Jonas Templestein, former CTO of Monzo Bank. Alongside my teammates, Hafsa Khan and Uzair Ahmed, we embarked on an ambitious project: to create an MVP of a large language model (LLM) aimed at detecting and fixing security vulnerabilities within a challenging 32-hour timeframe. Our project, ArmouAI, embodies a forward-thinking vision for cybersecurity, addressing the critical need for security in AI-driven code development.

What it does

ArmouAI stands as an innovative LLM designed to detect security vulnerabilities in code and provide fixes, aiming to revolutionize cybersecurity practices. Aiming to as a final product combine the principles of Cloud-Adaptive Threat Simulation with an Immune System approach, our solution represents a scalable and ground-breaking initiative. It leverages advanced AI to scrutinize code for potential security flaws, offering a proactive, all-in-one cybersecurity framework that prioritizes security from the ground up.

How we built it

Our approach to building ArmouAI was multifaceted and collaborative. Abdel spearheaded the LLM development, Uzair focused on web development, and Hafsa took charge of data collection and processing. We utilized the 'starcoder' model from Hugging Faces, training and fine-tuning it with a unique dataset focused on C and C++ vulnerabilities sourced from GitHub. Our model training incorporated the "sentence_transformers" library and FAISS for deduplication, emphasizing efficiency and precision. This intricate process involved leveraging the transformers library for fine-tuning, employing libraries like peft for parameter efficiency, and utilizing trl for supervised fine-tuning, culminating in a model that not only detects but also fixes security vulnerabilities.

Challenges we ran into

Our journey was not without its challenges. The ambitious scope of our project required rapid learning and adaptation, especially in the complexities of LLM development and the intricacies of fine-tuning for security applications. We encountered technical hurdles, such as optimizing the model within the constraints of limited resources and navigating licensing issues with the open-platypus database. Additionally, the pressing time frame of the hackathon tested our ability to deliver a functional MVP, pushing us to prioritize essential features while maintaining our commitment to innovation.

Accomplishments that we're proud of

Despite the hurdles, our team achieved remarkable milestones. With no prior experience in LLMs, we successfully developed and hosted an MVP of an end-to-end model, showcasing our ability to learn quickly and adapt under pressure. Our project's recognition among the TOP 6 entries at the hackathon validates our innovative approach and the potential of ArmouAI to make a significant impact in cybersecurity.

What we learned

This hackathon was a profound learning experience, teaching us the importance of determination, collaboration, and innovation in tackling complex challenges. We gained invaluable insights into LLM development, data processing, and the broader implications of AI in cybersecurity. Our journey underscored the critical role of security in software development, inspiring us to continue exploring and advancing in this vital field.

What's next for ArmouAI

Looking ahead, we envision expanding ArmouAI's capabilities to offer comprehensive vulnerability assessments and customized threat simulations based on specific software vulnerabilities. By integrating advanced AI techniques, such as Generative Adversarial Networks (GANs), we aim to provide a holistic cybersecurity solution that not only detects but also actively defends against cyber threats. Our journey at the hackathon is just the beginning, and we are excited about the potential to transform our vision into reality, contributing to a safer digital future.

Closing Thoughts

we extend oru heartfelt gratitude to everyone involved in the University of Cambridge GenAI Hackathon, especially the organizers and judges, for providing us with this incredible opportunity to learn, grow, and innovate. Here's to the next chapter in the ArmouAI story. Stay tuned!