Fujitsu, the global technology company, has introduced two new AI trust technologies aimed at enhancing the reliability of conversational AI models. The first technology focuses on detecting hallucinations in conversational AI models, where the AI generates incorrect or unrelated responses. The second technology, developed in collaboration with Ben Gurion University, detects phishing site URLs implanted in AI responses through poisoning attacks.
By incorporating these new technologies, Fujitsu aims to provide users, both corporate and individual, with a tool to assess the reliability of conversational AI responses, promoting a more secure use of AI across various applications. Professor Yuval Elovici from Ben Gurion University lauded the significance of Fujitsu’s hallucination detection technology for establishing trusted conversational AI systems.
Fujitsu will integrate these trust technologies into its conversational AI core engine offered through the Fujitsu Kozuchi – Fujitsu AI Platform, granting users access to powerful AI and machine learning tools. The technology to detect hallucinations will be available in Japan from September 28, 2023, while the technology to detect phishing URLs will be available from October 2023. Fujitsu plans to roll out both technologies globally in the future.
Conversational AI is commonly used by businesses to extract information from pre-registered data and provide accurate responses. However, hallucinations can occur when the AI fails to accurately extract relevant data, resulting in unrelated or incorrect replies. To address this, Fujitsu’s technology identifies parts of sentences where hallucinations are likely to occur by breaking down the AI’s reply and identifying named entities. This approach significantly improves the accuracy of hallucination detection compared to existing methods.
In terms of detecting manipulated URLs in AI responses, Fujitsu’s technology plays a crucial role in mitigating attacks that inject malicious information into AI training data. By leveraging the tendency for hostile entities to attack specific AI models, the technology detects manipulated data by evaluating the difference in rationale across various AI models. It not only detects phishing URLs but also enhances the AI’s resistance against deliberate misjudgments.
Fujitsu’s commitment to the Sustainable Development Goals (SDGs) also drives its purpose of making the world more sustainable through innovation. As a trusted digital transformation partner, Fujitsu focuses on resolving global challenges using its expertise in computing, networks, AI, data & security, and converging technologies.
With the launch of these AI trust technologies, Fujitsu aims to ensure the reliability and security of conversational AI models, paving the way for their widespread adoption across industries. By providing users with the means to evaluate the trustworthiness of AI responses, Fujitsu contributes to the development of a more secure and efficient AI-driven future.
Source: Fujitsu Ltd