ArmourAI

Armour AI

AI Hackmasters Hafsa Khan Abdelbasit Lemamsha Uzair Ahmed

Begin our journey...

Join us in our approach

Chosen Domain

Solution

Idea

Web Application

Data Collection

Technologies

The Future

Implemenatation

Obstacles

Developing Our Idea

Research

Generative AI coding solutions tend to follow a 'functionality' first approach. This neglects security issues posed to an organisation - which arguably leads to organisations straying away from utilising AI for their software egineering solutions

+info

Our Solution

Deploying LLM and training it on a database of coding vulnerabilities and their solutions intially -> in a smalll range of programming languages. This was done to explore effectiveness and viable usage of this idea.

+info

Current competitors in the market

• CrowdStrike - Charlotte AI
• LexisNexis® ThreatMetrix
• NVIDIA Morpheus
• Broadcom - SymantecAI
• AMD Xilinx - PFP Cybersecurity

+info

Implementing our idea

Challenges encountered

Model training was challenging as it required multiple attempts to accurately tokenize the data. Our vision was too high for these 24 hours we had to really simplify the idea down to its core MvP.

The Future

Providing users with a complimentary software vulnerability assessment and a brief overall score, without further explanations, is particularly effective for marketing purposes.

Employs deep learning to tailor attacks according to specific software and its vulnerabilities, derived from standard attacks (essentially a GAN). The model then provides feedback on the success of the attack and its severity.

Thank you!

The Website Application

The programming languages used for the implementation of the website was HTML, CSS, and JavaScript. Originally, SquareSpace was used but we realised we couldn not add the API link to the webpage therefore we proceeded with making the website through Visual Studio Code and implemented the requiremed scripts on the website.

Data Collection

Centre for Strategic & International Studies

Analysed a dataset containing cyber attacks that occurred between 2006 and 2024 in order to extrapolate the most prevalent types of cyberattacks and their corresponding impacts. Then, we discovered a labelled dataset on GitHub, which influenced our decision to proceed with this method rather than the others in order to identify common vulnerabilities in code and solutions.

Technologies used

LLM & Hugging Face

The technologies that we implemented for the LLM model was using Python, and Pandas, Transformers Library, seaborn for visualisations. StarCoder on base model and trained on Open Platapus database and a database from Github that contains vulnearabilities in C and C++.