The gendarmerie develops many use cases around transcription, text or even computer vision.
The AI accelerates its breakthrough in the public service: Dinum for the State, Amiad for the armies and … Datalab of the ANFSI (Digital Agency of the Internal Security Forces) for the gendarmerie. The latter multiplies concrete use cases with the aim of gaining operational efficiency and improving the service rendered to users.
AI at the service of transcription
This is the most industrialized use in the gendarmerie. The Datalab has developed a Speech to Text technical brick based on the open source Whisper of Openai. Several projects automatically transcribe speaking into text. The most advanced is a word, a system that automatically transcribes the hearings of minor victims. The system responds to a judicial imperative, because French law imposes the written transcription of these testimonies. “French law is a written right, and therefore we must transcribe these hearings filmed for the magistrate”, explains Marc Boget, director of the digital and technological strategy of the national gendarmerie. Already deployed in family protection houses, this system makes it possible to considerably reduce the processing time, going from eight hours per hearing only two to four.
Other applications of this same technology are being finalized: the transcription of telephone listening, also subject to the legal written transcription obligation, the mobility report allowing gendarmes to dictate their activities so that they are automatically classified, or the vocal memo for criminal identification technicians on crime scenes. On the technical level, the system required specific adaptations: “We have reached a Whisper model with specific children’s voice data for the hearing part of minors and specific settings for other use cases”, specifies Adrien Ly, chief of the Anfsi.
LLM, for the search for qualified information
Beyond the vocal transcription, the gendarmerie uses, more conventionally, LLM to optimize the search for information and simplify procedures, both for agents and for citizens. In this sense, the I-ACCUEIL project allows gendarmes responsible for welcoming the questions of citizens on various subjects. “When you are responsible for a brigade, you are asked by our fellow citizens on almost all the subjects,” explains Marc Boget. A chatbot, currently based on the LLAMA 3 70B model with a RAG system, indexes legal codes and professional documentation to provide specific responses. “The gendarme asks him a question in natural language, and the AI answers by tagging the origin of the information so that it can be verified,” explains the specialist. Currently in experimentation in a department, its national deployment awaits the acquisition of more computing power.
In another genre, the prednatinf system helps gendarmes to navigate in the complexity of the penal code by helping to categorize the offenses. “Each offense must be categorized, give it a code of a violation nature. There are more than 15,000,” recalls Marc Boget. The system allows agents to save precious time. “You type the summary of the facts in natural language and the system presents you what seems wise and coherent with what you have informed,” describes the general. A technology that should also benefit citizens by 2026 with the evolution of the online complaint, avoiding them “the complexity of the penal code” to choose the nature of the offense to declare.
An even more ambitious project concerns the “automatic synthesis of procedures”. The system aims to create summaries of legal proceedings to facilitate their transfer between investigators. “When you pass an investigator procedure A to an investigator B, you have an incompressible time which is the time required for investigator B to understand what is in the procedure,” recalls Marc Boget. Although the development is underway, challenges remain in matters of precision: “We are talking about legal proceedings and what is written must be fair. The tests have shown that AI still approximates certain things, which is not possible in a legal framework where the accuracy is essential.” The project is therefore still in development.
Computer vision for defined cases
Finally third and last pole of use cases around AI: the Computer vision. The automatic categorization of child pound images is a precious tool for specialized investigators. “When you go to a child psychoam author, you recover 100,000, 150,000, 200,000 images,” explains Marc Boget. Faced with this mass of data, the AI allows to sort and prioritize the analysis: “The tool allows to say, this first package, I am 99% sure that they are photos of landscapes. Conversely, these, I am sure that it is child pornography”. The system, already deployed, relieves investigators from the National Center for Analysis of PEOPORNOGRAPHICAL IMAGES of a task “particularly trying at the cognitive level”, while making it possible to process files within the constrained deadlines for police custody.
If this first application focuses on the analysis of static images, the gendarmerie is now gradually extends its expertise in video processing. A particularly promising project concerns the analysis of the videos captured during the research of missing persons. The gendarmerie helicopters regularly fly over large areas during research operations, but so far, efficiency depended only on human observation.
The principle is simple but effective: during research missions, the helicopters of the gendarmerie equipped with high definition cameras continuously film the areas overflown. “I am in a helicopter. I have a camera on the optronic ball. I film and I am looking for someone who has disappeared,” describes Marc Boget. Once the mission is complete, the videos recorded are processed, on the ground, by a visual artificial intelligence system capable of detecting human forms or anomalies that the crew would not have noticed during the flight. AI analyzes each image to locate any suspicious silhouettes or body and signals them. A delayed treatment which respects the current legal framework which does not yet allow video analysis in real time outside the specific context of the Olympic Games.
To maintain and develop its expertise in AI, the institution invests massively in continuing education, which represents half of the operating budget of the ANFSI. Training, targeted and pragmatic, with the best experts in the four corners of the globe, to allow teams to stay at the forefront. Faced with competition from the private sector, capable of offering remuneration up to ten times higher, the gendarmerie has developed a recruitment and loyalty strategy based on the sense of clearly defined missions and career paths. “If we do not give visibility and readability on their trajectory, we lose them,” said General Boget.