AI trust and Fujitsu's AI trust technologies in Conversational Generative AI
December 21, 2023
Japanese
Generative AI is opening up new possibilities for innovation in society and business. Various types of AI, including generative AI, are enabling companies to promote business growth and provide fulfilling customer experiences. Meanwhile, trust in AI (the trust that people and society have in AI) and AI trust (AI ethics, security, and quality) are critical to the spread of AI in businesses and society.
Fujitsu has long been committed to building trust in AI and contributing to sustainable development for society. One example is the Fujitsu AI Trust Research Center, which focuses on the research, development, and implementation of AI ethics, and the creation of tools such as the AI Ethics Impact Assessment and Fujitsu AI Ethics for Fairness. Now, the AI Trust Research Center has announced its hallucination detection and phishing URL detection technologies to realize trusted conversational generative AI that can be applied to corporate operations. Such technologies are further discussed in the whitepaper.
This document first examines the broad impact of AI trust on society at large, using the STEEP (Social, Technological, Economic, Environmental, and Political factors) framework. We also addressed AI trust in the enterprise because it is an important issue for enterprises seeking to leverage generative AI to improve competitiveness, increase sales, and reduce costs. The document then describes Fujitsu's activities to achieve trust in AI and its AI trust technologies.
Hallucination detection technology and phishing URL detection technology
The first technology detects hallucinations, in which the conversational generative AI returns answers that are not based on facts or established knowledge. We have observed that hallucinations tend to occur in the form of proper nouns and numerical values, and contents of replies tend to differ when questions are repeated. Based on these observations, our technology detects hallucination more precisely than existing methods by analyzing the semantics of the answers of conversational generative AI and identifying and focusing on the parts of proper nouns, numerical values, and other unique expressions. Fujitsu benchmarked this technology using open data, including the WikiBio GPT-3 Hallucination Dataset
(*1) and found that it could improve the accuracy of detection (AUC-ROC) (*2) by approximately 22% compared to other state-of-the-art methods for detecting AI hallucinations.
The second technology addresses the problem of conversational generative AI creating responses that include phishing URLs. In addition to identifying phishing URLs, Fujitsu’s new technology increases the AI’s resistance against existing attacks tricking AI models into making a deliberate misjudgment to ensure highly reliable responses by the AI. It utilizes technology jointly developed at the Fujitsu Small Research Lab(*3), which was established at Ben-Gurion University of the Negev(*4).
Comment from developers
AI Trust Research Center development team member
Satoshi Munakata
Taku Fukui
Vikas Pahuja
Comment from developers of hallucination detection technology:
Our project focuses on developing hallucination detection technology for conversational generative AI, with the aim of increasing its trustworthiness. This technology is key to ensuring that our customers can confidently use these AI systems in their digital transformation efforts, overcoming concerns about AI reliability and promoting safer, more effective AI integration.
Comment from developers of phishing URL detection technology:
In the realm of AI, generative AI stands out as a transformative force. Our dedicated research ensures it's not just powerful but also trustworthy. Consider Fujitsu's 'Phishing URL technology' - an innovation that shields users from deception during interactions with conversational AI. We're dedicated to creating secure technologies, ensuring our customers feel confident in using the advancements we bring for a safe AI experience. Your trust in our innovations is our top priority.
Future plan
The hallucination detection technology is already available to enterprise users in the Fujitsu Kozuchi (code name) - Fujitsu AI Platform's conversational generative AI core engine, and the URL detection technology will soon be available in the same core engine. We will first offer this service in the Japanese market and plan to expand globally in the future.
As AI is widely used in society, Fujitsu believes it is important to improve AI trust in terms of ethics, security, and quality. Therefore, we will continue research and development of technologies to improve AI trust, and work toward its practical application.
- (*1)WikiBio GPT-3 Hallucination Dataset
https://huggingface.co/datasets/potsawee/wiki_bio_gpt3_hallucination
- (*2)AUC-ROC (Area Under the Curve of the Receiver Operating Characteristic Curve):
The area under the curve of the curve obtained when the threshold value of the judgment is changed with respect to the abnormality score by placing the true positive rate on the vertical axis and the false positive rate on the horizontal axis. A random anomaly score is 0.5, and a perfect answer is 1.0. It is generally considered that a certain level of performance can be achieved when it is higher than 0.7. - (*3)Fujitsu Small Research Lab:
An initiative in which Fujitsu researchers reside or stay at the university on a long-term basis to accelerate joint research, discover new themes, develop human resources, and build medium- to long-term relationships with the university.
https://www.fujitsu.com/global/about/research/srl/ - (*4)Ben-Gurion University of the Negev:
Headquarters: Beersheba, Israel; President: Daniel Chamovitz
Inquiries regarding this matter
Share
Recommend
Connect with Fujitsu Research
