FSS deploys AI to catch crypto market manipulation, auto-analyze trading rooms

The Financial Supervisory Service is automating its surveillance of unfair trading in the virtual asset market using generative AI and machine learning. The system will flag suspected market manipulation and wash trading from real-time transaction data, and analyze signs of unfair trading found in online trading rooms, posts and videos. Once suspicious activity is detected, AI will support the process from trade analysis through to drafting review reports.

The FSS said Thursday it had built an AI market surveillance framework combining publicly available large language model-based generative AI with machine learning algorithms to strengthen its virtual asset market oversight. The system was designed and built entirely by in-house FSS staff.

The FSS had earlier developed an algorithm in January to automatically identify the asking price and market price involvement periods of suspected market manipulators — work that investigators had previously done by hand. In April, it added a function to automatically detect multiple linked accounts used in market manipulation. The latest expansion extends AI's role into market surveillance, including detecting and analyzing abnormal trading and assessing whether a formal investigation is warranted.

The new system uses real-time data from exchanges to flag coins suspected of ultra-short-term price manipulation. It identifies key manipulation patterns such as sharp price swings in specific time windows or in coins with restricted deposits and withdrawals. Generative AI then analyzes exchange notices and news related to the flagged coins to examine the causes behind sudden price and volume movements.

Benford's Law and machine learning algorithms will be applied to detect wash trading and matched orders. The system analyzes trading volume distribution and execution intensity by coin to screen out those deviating from normal patterns, then narrows down candidates to coins and periods where actual trading patterns diverge significantly from learned normal behavior. It also weighs price and volume fluctuations together to assess whether a formal investigation is needed.

Generative AI will also be used to monitor unfair trading activity online. The system collects the latest virtual asset-related posts and videos, converts video subtitles and audio into text, and analyzes the content for signs of illegal conduct. Targets include illegal front-running through online trading rooms, the spread of false information via video, and posts inciting unfair trading. Generative AI evaluates the risk level of each item by determining whether it constitutes trading inducement, a tip-off about another party's unfair trading, or simply casual chatter.

The deep-dive analysis that follows the detection of abnormal trading will also be automated. Suspicious accounts are screened based on metrics such as asking price involvement rate, trading profit and market price involvement rate, and algorithms are used to analyze suspect groups and suspect periods. Generative AI then automatically drafts a review report in a set format based on the calculated suspicion indicators.

However, the decision on whether to proceed with further analysis or a targeted investigation rests with the investigator in charge. The investigator will determine follow-up action based on the AI-drafted review report, and the reports will be stored cumulatively for use in future investigations.

The FSS plans to further strengthen its AI-based virtual asset market surveillance and investigation framework, including by developing additional functions to support fund-flow tracking and on-chain tracing.

Meanwhile, the FSS launched its integrated financial security control system "FIRST" in February. Building on that, it plans to pursue AI security threat response, broader AI governance and stronger IT internal controls going forward.