Use the neural-chat-7b Model for Advanced Fraud Detection: An AI-Driven Approach in Cybersecurity

Get the Latest on All Things CODE



neural-chat-7b and Bilic's Pioneering Customization

With the thriving presence of AI in our lives, Bilic, a cybersecurity company committed to using AI for financial security and compliance, has developed a fraud detection system based on the neural-chat-7b model from Intel. The goal is to improve existing fraud detection as a model for future applications that use AI technology.

Bilic's Vision for AI-Driven Financial Security

Traditional fraud detection methods rely on rule-based approaches, which are inherently constrained when confronted with the dynamic and sophisticated nature of contemporary fraudulent tactics. By incorporating AI into a contextually optimized dataset, we can enhance our ability to detect and comprehend fraudulent interactions in the application.

The Need for Contextual Understanding in Fraud Detection

Bilic prioritizes fraud detection because traditional methods are ineffective against modern tactics. Rule-based approaches used in the past are limited when it comes to dealing with the ever-evolving and complex nature of fraudulent activities today.

The Journey for Fine-tuning neural-chat-7b for Fraud Detection

Craft a Custom Dataset

Bilic focused on developing a distinct dataset for fraud detection. The team carefully designed this dataset to mirror real-world scenarios that represented the diverse spectrum of fraudulent activities and paired them with summaries that encapsulated the essence of these interactions.

Examples of the data used included various fraudulent conversations via chat or email. Within this dataset, each entry comprises a dialogue alongside a corresponding summary, thoughtfully structured to guide the model in summarizing and evaluating the conversation for potential fraudulent content. This dataset plays a pivotal role in fine-tuning the neural-chat-7b model, and enabling it to discern and identify the nuances inherent in fraudulent dialogues.

Fine-tuning Methodologies

Bilic adopted a comprehensive and multifaceted approach for fine-tuning:

Model Preparation

The first step involved loading the neural-chat-7b model with configurations tailored to the fraud detection dataset. This preparation phase was crucial in adapting the model to fraudulent conversations' unique nuances and patterns.

Parameter Adjustment Using LoraConfig

We used LoraConfig to control training parameters precisely. This low-rank adaptation technique allowed us to modify only a fraction of the model's parameters, making the fine-tuning process efficient and effective.

Progressive Embedding Fine-Tuning (PEFT) and Gradient Checkpointing

To efficiently adapt the model, we implemented PEFT, specifically focusing on embedding layers with LoRA (Low-Rank Adaptation). Alongside, gradient checkpointing was enabled to optimize memory usage during the training of this large-scale model.

4-bit Quantization with GPTQConfig

A pivotal step in our process was applying 4-bit quantization. This technique, achieved using Intel® Neural Compressor, supports the GPU post-training quantization (GPTQ) algorithm, which adapts to more large language models (LLM), significantly reducing the model's size and computational requirements without compromising its performance. This was particularly crucial for deploying the model efficiently in real-time fraud detection scenarios.

Strategic Training Management Using TrainingArguments

The training lifecycle was managed using TrainingArguments from the Hugging Face* library. This involved fine-tuning elements such as batch size, learning rate, and learning rate scheduler to optimize the training process. Such strategic management was key to balancing efficiency with the model's effectiveness.

Achievements and Outcomes

This fine-tuning resulted in a model that not only identifies potential fraud but also provides comprehensive summaries of conversations, offering insights into why a particular interaction might be deemed suspicious. This capability significantly enhances the model's utility in advising and protecting potential victims.

The Future of AI in Financial Fraud Detection and Bilic's Role

The integration of AI into the financial security domain is evolving from rule-based systems toward more dynamic, contextually aware, and adaptable models. Bilic's efforts exemplify how AI can be used to safeguard financial transactions and ensure compliance with regulatory mandates within an ever-expanding digital landscape.

"Our collaboration with Intel, especially in utilizing the neural-chat-7b model and Intel® Liftoff, symbolizes a groundbreaking chapter for Bilic. This joint effort is not just about integrating an advanced AI model; it is about co-creating a robust, intelligent system for fraud detection. Intel Liftoff has played a crucial role in enabling us to seamlessly deploy and scale our solutions, ensuring that our systems are as agile and adaptive as the threats they are designed to counter. Together with Intel, we are forging a new frontier in AI-driven financial security, setting a new benchmark in the industry."

— Saminu Salisu, CEO of Bilic

"Intel is working to bring AI everywhere. We are glad to see that the neural chat model from Intel has been used by Bilic in advancing fraud detection for financial security, a great example of harnessing AI technology to improve complex analysis."

— Huma Abidi, general manager and senior director of AI Software Products and Engineering, Intel

As we advance, Bilic's visionary application, in partnership with Intel's technological expertise, establishes a benchmark in the use of AI for fraud detection. Without a doubt, the insights gained, and the methodologies crafted by Bilic will exert a substantial influence on forthcoming advancements in AI-driven financial security and regulatory compliance.

What’s Next?

Read this article on how the neural-chat-7b model has achieved the top ranking for seven-billion-parameter models on the Hugging Face Open LLM Leaderboard. This highlights how neural-chat-7b achieves comparable accuracy scores to models 2-3x larger than it.

We encourage you to also check out and incorporate Intel’s other AI and machine learning framework optimizations and end-to-end portfolio of tools into your AI workflow and learn about the unified, open, standards-based oneAPI programming model that forms the foundation of the Intel® AI Software Portfolio to help you prepare, build, deploy, and scale your AI solutions.