GenAI Guardian — Project Overview

1. What Is GenAI Guardian

GenAI Guardian is a special fraud detection edge AI system that operates completely offline.

No internet connection is required whatsoever. Voice and video data never leave the device.

Elderly individuals receive phone calls at home — telecom fraud strikes in this most private of spaces. If a system introduces cameras and microphones into that space while transmitting data externally, neither the user nor their family can feel at ease. No matter how much one explains that "the data is encrypted" or "the server is secure," anxiety persists as long as a communication pathway exists. GenAI Guardian resolves this problem through a design that simply does not require any communication pathway. Furthermore, the system is linked to the telephone and activates only when a call is in progress. There is no constant surveillance.

Design Principle	Details
Communication	Not required. Fully offline operation
Data Transmission	None. Voice and video data never leave the device
Activation	Phone-linked. Active only during calls (no constant surveillance)
Hardware	Compact camera (with microphone and speaker) + dedicated compact device
Privacy	Non-intrusive. Does not record or accumulate data from the user's daily life

For deployment in elderly households, neither the availability of an internet connection nor communication costs pose a barrier.

By fusing insights from criminal psychology with AI technology, the system detects the "structure of psychological manipulation" used by fraud perpetrators in real time, through both the conversational context of call audio and behavioral patterns captured on camera.

Pillar	Function	Characteristics
Context Token Detection	Real-time contextual analysis of call audio	Purpose-built for fraud prevention. Context-aware
Human Behavior Analysis	Detection of behavioral changes under duress	On-device AI-based motion recognition
Data Foundation	Multilingual, multi-regional anti-fraud data	Approx. 600 hours of audio, approx. 4 billion characters of text

2. The Social Problem: Escalating Telecom Fraud

The Scale of the Problem — China and Japan

GenAI Guardian was developed and validated in two of Asia's most affected markets.

Market	Cases (2024)	Annual Losses (2024)	Source
China	294,000 cases	¥97 billion (~$13.5B)	Ministry of Public Security
Japan	21,043 cases	¥71.8 billion (~$480M)	National Police Agency

In China, telecom fraud accounts for roughly 60% of all criminal cases. Despite intercepting 4.69 billion scam calls and blocking ¥315.1 billion in fraudulent transactions in 2024 alone, direct losses still reached ¥97 billion. In Japan, fraud damages nearly doubled in 2025, reaching an all-time high of ¥141.4 billion.

A Growing Global Threat

This problem is not limited to Japan. In January 2026, U.S. anti-fraud expert Frank McKenna stated in his annual report (2026 Fraud Predictions) that AI-driven fraud has entered an "era without guardrails."

Global fraud losses exceeded $1 trillion in 2025
Deepfake attacks increased 3,000% year-over-year
Approximately 70% of adults worldwide have experienced some form of fraud
AI automation is giving rise to "fraud factories" capable of targeting tens of thousands of people daily with minimal staff

The evolution of AI is making fraud techniques more sophisticated by the day. Countermeasures must also evolve through AI.

The Structure of Elderly Victimization

In Japan, 65.4% of telecom fraud victims are aged 65 or older (National Police Agency, 2024). In China, over 70% of impersonation fraud victims are elderly. Both countries face rapidly aging populations: Japan's 65+ population has reached 29.3% (the world's highest), while China's 60+ population surpassed 310 million in 2024 (22%) and is projected to exceed 400 million by 2035.

The structural risk of telecom fraud is expanding in lockstep with demographic aging — and this is not limited to Asia. The FBI reported $4.9 billion in losses among Americans aged 60+ in 2024 alone.

Image depicting telecom fraud

Limitations of Existing Countermeasures

Current primary countermeasures — awareness campaigns, nuisance call filters, and call recording — either intervene only after a call has already begun or rely on the victim's own judgment.

However, fraud perpetrators succeed precisely by overriding victims' judgment. If a victim were capable of thinking "this might be a scam," the fraud would not succeed in the first place. What is needed is a system capable of detecting when a person has lost the capacity for sound judgment.

3. Established Institutional Attention to This Challenge

Anti-Fraud Measures Through AI and Criminal Psychology

Anti-fraud approaches that fuse AI with criminal psychology have attracted growing attention internationally.

Fujitsu Limited, Toyo University, and the city of Amagasaki in Hyogo Prefecture have been conducting joint research since 2022, pursuing telecom fraud prevention through "converging technology." In May 2025, they achieved a fraud detection accuracy of 82%, which was covered by major Japanese media outlets.

Date	Milestone
March 2022	Joint research launched. First field trial (Amagasaki City Hall)
September 2022	Identified 11 factors related to victims' psychological states. Presented at the Japanese Association of Applied Psychology
October 2023	"Converging Technology Research Group" adopted at the Japanese Psychological Association
November 2023	Developed a generative AI-based telecom fraud training tool. Held experience sessions in Amagasaki
November 2024	Deployed devices in 22 elderly households for validation in daily living environments
May 2025	Achieved fraud detection accuracy of 82%

This initiative has been officially communicated through multiple channels — Fujitsu, universities, municipalities, and national newspapers.

Relevance to GenAI Guardian

The Toyo University professor leading this research was the academic supervisor of the GenAI Guardian researcher during their master's program. The researcher has also participated in the Converging Technology Research Group, attending research presentations and contributing to discussions.

GenAI Guardian addresses the same social challenge through a different approach.

Comparison	Fujitsu Joint Research	GenAI Guardian
Detection Method	Millimeter-wave sensors + physiological responses	Context Token Detection + Human Behavior Analysis
Focus	Hardware sensing	Software / Edge AI
Device	Dedicated sensor installation	General-purpose camera + compact device
Communication	—	Fully offline
Shared Philosophy	Non-intrusive, privacy-preserving	Same

As different research trajectories toward the same goal, there is potential for mutual complementarity.

4. Approach

4.1 Context-Based Token Detection — How It Differs from Conventional Methods

The system analyzes call audio in real time to assess the likelihood of fraud.

Conventional keyword-based systems trigger an alert when specific words such as "money," "transfer," or "police" appear. However, these words are also used frequently in everyday conversation. The result is a flood of false positives, causing users to ignore alerts altogether.

GenAI Guardian takes a fundamentally different approach.

Rather than individual words, it reads the entire flow of conversation as context
It determines whether a pattern of psychological manipulation is present in the conversation — the elicitation of fear, the creation of time pressure, the inducement of social isolation
It does not react merely because a specific word appears. Detection occurs only when those words are embedded within a manipulative context

This context awareness suppresses the false positive problem at a fundamental level.

4.2 Human Behavior Analysis Under Duress — Why It Is Especially Effective for the Elderly

On-device AI analysis of camera footage detects behavioral changes in the victim during a call.

It is well established that when humans are psychologically cornered — experiencing fear, anxiety, or confusion — they exhibit specific behavioral patterns: restless movements, repetitive actions, unusual changes in posture, and so on.

This detection is particularly effective for elderly individuals. Because the range of movement variation in daily life is smaller for the elderly compared to younger people, any deviation stands out more sharply, inherently increasing detection reliability.

All video analysis is processed by on-device AI, and no video data is transmitted externally.

5. Researcher Background

Practical Experience in Criminal Courts

Prior to coming to Japan, the researcher spent over two years working in criminal courts in China.

During this period, the researcher was involved in the trial and sentencing of a large-scale telecom fraud case that shook the entire country and extended into Southeast Asia. While a typical criminal verdict runs 4 to 6 pages, the verdict for this case exceeded 30 pages.

All information handled in criminal courts is subject to strict confidentiality obligations, with the exception of the final verdict. Through this experience, the researcher gained an understanding of the psychology, methods, and organizational structures of fraud from within the judicial process.

Academic Research

In 2021, the researcher's master's thesis focused on "The Relationship Between Anxiety and Vulnerability to Fraudulent Business Practices and Scams." From the perspective of criminal psychology, the research analyzed the structure of human vulnerability to fraud victimization.

International Experience

The researcher has lived and worked abroad for over 10 years, with trilingual (Japanese, Chinese, English) research and professional capability. This long-term immersion has cultivated a deep understanding of social structures, aging realities, and fraud victimization patterns across both China and Japan.

6. Data Assets

Aggregate Metrics

Metric	Value
Number of Datasets	30
Total Recorded Audio	Approx. 600 hours (equivalent to about 1 month of continuous listening)
Total Text Volume	Approx. 4 billion+ characters (equivalent to approx. 40,000 books)
Languages Covered	Japanese, Chinese, English, Korean (4 languages)
Geographic Coverage	Japan, China, United States, South Korea, International (5 countries/regions)
Data Period	2015 – 2025 (approx. 10 years; core period: 2022–2025)

Dataset Inventory by Category

#	Source Type	Region	Language	Scale	Summary
1	Joint research with a major telecom carrier	China	Chinese	200+ hours + hundreds of millions of characters	Call audio and transcribed text
2	Same as above (text-focused)	China	Chinese	Tens of millions of characters	Text classification data
3	English-language fraud call recordings (research use)	U.S./International	English	200+ hours + hundreds of millions of characters	Call recordings with research labels
4	U.S. federal government agency	U.S.	English	Tens of hours	Public record data
5	Major semiconductor company	International	English	Tens of hours	Voice evaluation data
6	Actual correspondence with fraud perpetrators	International	English	Hundreds of messages / tens of millions of characters	Primary source material accumulated over years
7	Japanese law enforcement agencies (5 agencies)	Various regions in Japan	Japanese	Several hours + tens of thousands of characters	Multi-format data from law enforcement sources
8	AI safety research (4 major tech companies)	International	English	2 billion+ characters	Safety evaluation data
9	South Korean academic research	South Korea	Korean	Hundreds of millions of characters	Academic research data
10	Voice security research	China/International	Chinese/English	Hundreds of millions of characters	Voice security research data
11	Deepfake detection	International	English	Millions of characters	Detection research materials

These datasets were systematically collected through years of investigation and relationship-building with law enforcement agencies and research institutions across multiple countries. Some data was obtained through investigative activities that involved personal risk.

Detailed information about the data assets can be presented in a controlled environment upon inquiry.

Data Statistics Dashboard

GenAI Guardian Data Asset Statistics

Data Type Statistics

File type analysis conducted on the same workstation. Audio data: 36,962 files / 33.5 GB. Text data: 57,428 files / 9.1 GB. Total: 108,105 files / 55.5 GB.

7. Future Outlook

GenAI Guardian's starting point is protecting the elderly from telecom fraud.

However, the system's detection target is not a specific set of fraud techniques but rather the structure of psychological manipulation through language itself. Tokens are the smallest building blocks of both human language and AI language processing — whether text is generated by AI or spoken by a human, it is all processed as sequences of tokens.

By its very nature, context-based Token detection can extend its scope of application in the following directions.

Phase	Target
Present	Detection of elderly telecom fraud (validated in China and Japan)
Near-term	Safety screening of AI-generated content; trust verification for digital humans and live streaming
Long-term	Multilingual expansion. As AI becomes more involved in communication, the applicability of context Token detection expands accordingly

8. Contact

Zichao Zhou [email protected]

GenAI Guardian — Project Overview 2026