Olfeo and AI: working together to classify content

As part of France's "AI Action Week," which took place from February 6 to 11, it is essential to highlight the fundamental role of artificial intelligence in cybersecurity. Far from being a mere trend that emerged with ChatGPT, AI has long been a lever for optimization and performance in the detection and prevention of cyberattacks.

At Olfeo, we have integrated these advanced technologies to improve the ranking and categorization of domains and URLs, thereby contributing to enhanced security.

A single database

With over 20 million domains and hundreds of millions of URLs listed, Olfeo stands out as the only French company to offer a database of websites that is finely classified according to usage and the French legal framework. This database is the nerve center of our solution, guaranteeing unparalleled accuracy and relevance in website classification, while limiting false positives.

The quality and comprehensiveness of our database are essential to ensuring effective cybersecurity.

Until a few years ago, classification was done manually by human operators, which ensured a high degree of accuracy but could obviously pose scaling issues.

With the emergence of AI models around ten years ago, an initial level of automation was introduced, but it was still supervised and confirmed by a human operator: pre-classification based on semantic analysis and Bayesian inference to deduce sets of category probabilities. For example, a site could be classified as "travel" with a 70% probability, "photography" with a 20% probability, or "blog" with a 10% probability. It was up to the human operator to validate the correct suggestion. This approach accelerated the classification process while ensuring impeccable quality.

However, faced with the exponential growth in the number of new domains created each month, this strategy needed to be supplemented by innovative technologies in order to maintain its accuracy and effectiveness.

One of the major challenges in classifying URLs is identifying malicious sites. Every month, nearly 70 million new domains are created worldwide, of which around 13 million are for malicious purposes. But limiting ourselves to malicious sites alone means overlooking legitimate sites and the value of being able to categorize these sites in detail. In fact, detailed categorization, even of legitimate sites, provides insight into how employees use the internet, but also allows access and security policies to be applied. One example would be the detection and categorization of SaaS applications, also known as Shadow IT, which allows for better management of the operational and security risks associated with the use of SaaS applications.

Learning to classify large volumes of data

That is why, in order to take AI to the next level, Olfeo has partnered with experts from various fields, notably as part of the interdisciplinary RAPID METIS project.

The METIS project, with a budget of €1 million, co-financed by the RAPID (Dual Innovation Support Scheme) program and supported by the Defense Innovation Agency of the French Ministry of the Armed Forces and the Directorate General of Armament, was launched in 2020. It aimed to implement AI and machine learning algorithms to improve the performance of Olfeo's internal classification tools. The project includes a key innovation dimension around data classification and machine learning methods to identify the category to be associated with a URL based on analysis of the web page content.

Carried out in collaboration with the University of Reims Champagne-Ardenne and the computer science laboratory at the University of Grenoble Alpes over a period of three years, the project combined exploration of the state of the art in semantic analysis and classification, the implementation of automated processing procedures, training of various machine learning models with existing Olfeo data, and finally the deployment of a deep learning model thatoptimized content recognition rates (based on a variant of the BERT model).

The project fully met its objectives. The relevance of the classification was improved by up to 30% for certain content categories, while maintaining the quality of the existing database with a "fair" URL recognition rate of close to 100%. These results enabled the classification algorithm to be integrated into the Olfeo pre-qualification tool.

Following this success, Olfeo continued its dual approach, combining classification using machine learning techniques while allowing humans to retain control over the final classification. With the increased accuracy of automatic classification, Olfeo is paving the way for automatic validation of certain types of content in the near future.

On-the-fly classification of URL requests: dream or future reality?

With the increasing quality and constant innovation of LLMs (Deepseek, Mistral, OpenAI, etc.), it is legitimate to ask whether it will eventually be possible to categorize or classify a request on the fly, without checking a database. The advantages would be significant, particularly the possibility of no longer having to maintain and update URL or domain databases. Given the large number of domains created, this could be an effective solution (perhaps the only one that can be scaled up?).

However, this approach has several limitations: first, the importance of latency. On-the-fly classification requires analysis and the application of a classification model to each user request, which necessarily means longer request times. At present, even with powerful models, latency is a problem.

The other limitation is the quality of the results. As powerful as today's models are, it is not possible to guarantee a 100% correct result. Not without the intervention of a human operator. False positives are a challenge for IT teams, forcing them to investigate and spend time validating or invalidating the result.

For the time being, the best approach is still to involve a human operator to keep the result databases as clean as possible.

Conclusion

The integration of AI into cybersecurity is no longer just an option, but a necessity. At Olfeo, we have made artificial intelligence a strategic asset to ensure accurate and effective website classification, while strengthening the security of digital infrastructures. By combining human expertise and artificial intelligence, we remain at the forefront of innovation to offer our users optimal protection against constantly evolving threats.

AI Action Week is an ideal opportunity to remind ourselves that AI is a major driver of transformation in many sectors, and that its role in cybersecurity continues to grow. Olfeo is fully committed to this dynamic in order to continue offering ever more powerful solutions that are tailored to the challenges of tomorrow.