Macnica and Elith conduct Japan's first proof-of-concept (PoC) for detecting fraudulent phone calls using Edge AI – confirming a detection accuracy of over 90% with real-time inference using Broadcom's NPU.

Macnica (Headquarters: Yokohama City, Kanagawa Prefecture, Representative Director and President: Kazumasa Hara, hereinafter referred to as Macnica) and Elith, Inc. (Headquarters: Bunkyo-ku, Tokyo, CEO: Koki Inoue) today announced that they have conducted the first PoC in Japan ^*1 to detect fraudulent phone calls using Edge AI. Utilizing a PON MAC SoC ^*3 evaluation board equipped with Broadcom's NPU (Neural Processing Unit) ^*2, they confirmed a detection accuracy of over 90% in a real-time inference evaluation environment built specifically for this PoC. This is expected to be useful in preventing fraudulent phone calls in fixed-line telephone environments.

■ Social Issues and the Background of Development
According to the National Police Agency, the number of reported special fraud cases in 2025 is projected to reach 27,832, with losses amounting to 142.31 billion yen, both record highs ^*4. Approximately 80% of special fraud cases begin with telephone contact, and among those aged 60 and over, many cases begin with landline telephone calls ^*5.
Furthermore, in July 2026, the Ministry of Internal Affairs and Communications requested telecommunications carriers to further strengthen measures to prevent damage from special frauds and other scams. The request also includes efforts to raise awareness and promote the use of AI-based fraud detection and warning services for landline users. ^*6​
Traditional methods of combating nuisance calls primarily relied on identifying phone numbers, but with the increasing sophistication of fraudulent tactics, there is a growing need for measures that can detect signs of fraud early on, based on the content and flow of conversations after the call is received. On the other hand, cloud-based voice analysis systems raise concerns about privacy due to the external transmission of voice data, as well as operational burdens such as server and cloud usage fees.
Therefore, Macnica and Elith focused on Edge AI ^*7, which performs AI inference on the device side, and conducted this Proof of Concept (PoC).

■ Overview and evaluation results of the PoC
This proof-of-concept (PoC) was realized through a joint effort between Macnica and Elith, covering everything from AI model development and performance evaluation to implementation verification in edge environments, utilizing Broadcom's NPU-equipped PON MAC SoC. By combining the AI technologies and communication infrastructure expertise of both companies, we are proceeding with verification toward the practical application of real-time fraud call detection technology.
Specifically, the call audio input via a USB speakerphone or similar device is analyzed in real time on the terminal side of an evaluation environment simulating a home gateway ^*8, and signs of a special fraud are determined from the conversation content and context (see Figure 1).

<Figure 1: Schematic Diagram of the Fraud Detection PoC Using Edge AI>

In this Proof of Concept (PoC), we utilized publicly available voice data to evaluate multiple special fraud scenarios, including "ore-ore" (impersonation) scams, scams impersonating government agencies, and call center-type scams, as well as general conversation scenarios. As a result, we confirmed a detection accuracy of over 90% in the evaluation environment built for this PoC. This demonstrates the potential of using NPU-powered Edge AI to determine fraud risk in real time during calls on communication devices.

Please note that these results are based on evaluations in a Proof of Concept (PoC) environment, and actual performance may vary depending on the usage environment, call quality, conversation content, and other conditions.

■Features of this solution
1. Real-time fraud detection using NPU/Edge AI
By performing inference processing using Edge AI on the terminal side, such as the home gateway where the landline phone connects, rather than in the cloud, call content is analyzed in real time to detect signs of fraudulent calls. High-efficiency AI inference is achieved by utilizing NPU processing.

2. Privacy protection for audio data
While cloud-based AI requires sending voice data to an external server, this solution completes AI processing within the device itself. This contributes to protecting user privacy and reducing the risk of information leaks.

3. Low-cost operation with reduced cloud usage fees.
By performing AI inference processing at the edge, server usage fees and cloud operation costs can be reduced compared to configurations where voice analysis is performed in the cloud. This makes it easier for telecommunications carriers and service providers to implement.

4. Proof of Concept (PoC) for application in a fixed-line telephone environment
We investigated the feasibility of a fraud detection technology that uses Edge AI to analyze the context of conversations in real time, assuming a combination of a home gateway and a landline phone. This technology is expected to complement conventional measures that use phone number information and be applied to new measures that can detect signs of fraud from the content of conversations after a call is received.

■Future Outlook
Based on the knowledge gained from this Proof of Concept, Macnica and Elith will work to expand evaluation scenarios and improve the performance of AI models, and will also conduct verifications that simulate actual usage environments through collaboration with telecommunications carriers and telecommunications equipment manufacturers.
Furthermore, we will continue to examine the feasibility of implementation in communication devices, including home gateways, and the operational challenges involved. Our aim is to contribute to the realization of a safer and more secure communication environment through its use in combating special fraud.

*1: Based on Macnica Elith's research (as of September 2026) on Edge AI type services or PoC cases that detect special fraud within Home Gateways or landline phones, as publicly available in Japan.

*2: NPU (Neural Processing Unit)
This processor is ideal for inference processing using pre-trained AI models. It efficiently executes AI calculations and supports high-speed, power-efficient inference processing on the terminal side.

*3: PON MAC SoC (Passive Optical Network Media Access Control System-on-Chip)
This is a SoC (System-on-a-Chip) that integrates communication control functions (MACs) for optical access networks (PONs) and other features onto a single semiconductor chip. It is used in communication equipment such as home gateways and optical network terminals.

*4: National Police Agency, "Report on the Recognition and Arrest Status of Special Fraud and SNS-Type Investment/Romance Fraud in 2025 (Final Figures)"
https://www.npa.go.jp/bureau/safetylife/sos47/new-topics/260605/01.html
https://www.npa.go.jp/bureau/safetylife/sos47/assets/img/new-topics/detail/260605/01/01.pdf

*5: National Police Agency, "Report on the Recognition and Arrest Status of Special Fraud and SNS-Type Investment/Romance Fraud in 2025 (Final Figures)"
https://www.npa.go.jp/bureau/safetylife/sos47/assets/img/new-topics/detail/260605/01/02.pdf

*6: Ministry of Internal Affairs and Communications, "Request for further strengthening of measures to prevent damage from special fraud, etc." (July 29, 2026)
https://www.soumu.go.jp/menu_news/s-news/01kiban18_01000288.html
https://www.soumu.go.jp/main_content/001084332.pdf

*7: Edge AI
This technology executes AI processing on devices closer to where the data is generated, such as terminals and communication equipment, rather than on cloud servers. It is expected to improve real-time performance, reduce data traffic, and enhance privacy.

*8: Home gateway: A device that combines multiple functions (ONU, router, telephone adapter, etc.) provided by the internet service provider when using the internet or Hikari Denwa (fiber optic telephone) via a fiber optic line into a single unit.

[For product inquiries, please contact us here]
Macnica FINESSE Company Company, Inc. - Ohara/Sakai
TEL: 045-470-9831
E-mail: bcm_wifi_support@macnica.co.jp

Elith Co., Ltd. Sawa Fubuki
TEL: 090-6602-9320
E-mail: fubuki.sawa@elith.ai

*Company names and product names mentioned in this text are trademarks or registered trademarks of their respective companies.
*Broadcom and the Broadcom logo are trademarks of Broadcom Inc. and/or its affiliates. Other company and product names mentioned are trademarks or registered trademarks of their respective companies.
*The information contained in this news release (including product prices and specifications) is current as of the date of publication. Please be aware that this information may change without prior notice.