Define generative artificial intelligence by most measures, and there is broad consensus: it’s transformative, with potential for both benefits and risks. The former has dominated discourse in recent years, first with assistive chatbots like ChatGPT, and later with models that generate images, video, and more.
Yet this attention has often overshadowed the need to address the latter. As generative AI advances, deepfakes have become more sophisticated. Impersonation is now cheaper to attempt and harder to defend against, raising risks in financial services where verifying identity is critical. Face swaps can bypass selfie checks, identity theft can slip through onboarding systems, and forged documents can evade manual review. The takeaway is straightforward: if detection lags, financial losses rise and trust erodes.
Defending against malicious deepfakes is not just about spotting a blurred edge or an odd pixel. Attackers constantly devise new tricks, while detection systems often rely on patterns that become outdated quickly. Public datasets are limited, which means models risk overfitting to what they know and failing to catch what they do not. Single-signal approaches, such as analyzing only images, leave blind spots. Strong defenses must work across varied lighting, devices, faces, and forgery styles. They also need to pair visual checks with other signals when images alone cannot provide certainty.
This challenge shaped this year’s “FinVolution Global Data Science Competition,” which focused on visual deepfake detection.
On September 24 in Shanghai, the competition’s grand final gathered ten finalist teams from 426 teams and 652 participants worldwide. They presented algorithms designed to distinguish authentic images from manipulated ones under conditions resembling real-world business use. A judging panel of experts from Fudan University, Zhejiang University, and the Chinese Academy of Sciences evaluated the algorithms for model architectures, training regimens, and deployment practicality.
Cross-domain generalization as the decisive test
If last year’s competition focused on deepfake voice detection, 2025 has broadened the aperture to the visual domain, mirroring how generative techniques have become multimodal. Fraud tactics such as face swapping and identity theft have grown more sophisticated in turn. “As technology improves and large models proliferate, the cost and barrier of deepfakes keep falling, making fraud risks more visible,” said Chen Lei, vice president of FinVolution, in a post-event media interview on September 24. “In finance, where identity verification is stricter than in most sectors, that creates real challenges. No single company can tackle this alone, so we hope the competition can spark more awareness and collaboration.”
The contest simulated real-world conditions. Finalists trained and validated their algorithms on a dataset of more than 100,000 facial images spanning geography, ethnicity, lighting, and image quality, variables that resemble the messy data fintech players encounter, especially in overseas markets. A proprietary dataset added fakes created by recent face swapping techniques, raising the bar by forcing teams to grapple with unfamiliar artifacts.
The decisive twist was this: while participants were exposed to only a handful of known attack methods during development, they were evaluated on their ability to detect several unseen types in the final. That design made cross-domain recognition and generalization the central task.
Chen noted that algorithms can become “overly sensitive” if they are disproportionately trained on existing datasets. “If a model relies too much on past data, its predictive power for future, unseen attacks declines. Sometimes, intentionally limiting what a model sees can help it generalize better,” he said.
That dynamic was evident during the competition, as several teams observed that many face swap forgeries leave fingerprints in the high-frequency components of an image. By moving into the frequency domain, measuring subtle statistical differences in those features, and tailoring treatments accordingly, they reported meaningful gains.
Others rethought the problem setup, subdividing the notion of fakeness into clusters based on forgery methods rather than treating it as a single category, and enriching training data to simulate unseen threats.
The shared goal was to break the pattern of detection trailing generation.
“The participants’ explorations hold real-world application potential,” Chen said. “Through this competition, we aim to continue fostering innovation and integrating outstanding outcomes into business scenarios with the whole industry. We will keep iterating in areas such as dynamic video deepfake detection and cross-scenario generalization to build the financial defenses of the future.”
Turning competition results into solutions
The FinVolution Global Data Science Competition has run annually since 2016, bringing together nearly 10,000 participants over the past decade to tackle problems at the core of digital finance, including credit scoring, fraud detection, user behavior analysis, dialect recognition, and, more recently, deepfake detection. The ethos is consistent: surface new ideas through the competition, then refine them into practical solutions.
According to Chen, AI now underpins a broad range of functions at FinVolution, spanning customer acquisition, risk management, operations, and more:
- FinVolution engineers use AI-assisted coding, with roughly one-fifth of code now generated using AI with human oversight.
- In marketing, about half of its digital ad creatives are said to be AI-generated. The firm’s legal teams use drafting and review assistants for support.
- In customer service, agentic systems, autonomous software that can plan and take constrained actions toward a goal, help surface relevant user data and suggest responses in real time, while chatbots and voice bots assist with routine issues.
- On the risk front, voiceprint analytics help detect organized groups of borrowers attempting to evade debt collection by identifying recurring speakers across calls and combining those signals with speech-pattern clues.
Zeta, the company’s AI-powered innovation platform, ties much of this together. According to FinanceAsia, as of the end of 2024, Zeta supported more than 1,000 large model applications and integrated DeepSeek’s R1 model. The reported outcomes are concrete: advertising content production costs have reportedly fallen by about 60%, and voice summarization tools used for customer service use cases are said to have boosted agent productivity roughly twentyfold, significantly cutting call handling time.
In the second quarter, FinVolution reported that its defense stack against deepfake attacks reached 98.8% detection accuracy. Its in-house visual AI system reportedly identifies forged images with up to 95% accuracy. To further strengthen protection, the company has developed layered dynamic facial verification, randomized voice prompts, and real-time video authentication.
Crucially, FinVolution representatives stressed that competition outputs rarely translate into products directly. Instead, they seed directions that product teams can validate, harden, and deploy. “Competitions focus on maximizing scores, which differs from production requirements. But the ideas are valuable, such as cross-domain adaptation, which some teams explored this year, giving us new perspectives for our own research,” said Wang Chunping, chief AI scientist of FinVolution.
“It doesn’t have to be a full solution, as even small insights can be applied in practice. Often, it’s the fresh thinking that inspires us.”
Beyond systems, FinVolution is deepening ties across the research ecosystem. This year’s competition became an official partner of the International Joint Conference on Artificial Intelligence (IJCAI) and was included as an official track of the Conference on Information and Knowledge Management (CIKM), with outstanding teams eligible to share results internationally. These links aim to translate frontier methods into reliable, responsible AI for real-world finance.
AI driving FinVolution’s globalization
“Every country is different,” said Chen when asked about the firm’s international expansion.
The point is practical. Business and technical realities often diverge across borders, which, in his view, is why AI matters. In some markets, GPU access is scarce, pushing teams to compress or distill models and, when needed, deploy smaller architectures optimized for CPUs. Data gaps add another layer of difficulty. Credit bureau coverage tends to be lower, and identity systems are fragmented, with multiple ID types in circulation instead of a single national ID.
The result, Chen said, is a need for localized model variants and continual retraining, along with multimodal detection that looks beyond pixels to behavior and context. “In some countries, there are more than a dozen types of identity documents. That means our facial recognition system would actually need to adapt to handle each ID type, whereas in China, we only need to develop one algorithm for the national ID card,” he said.
“This may be hard to imagine in China, where everyone is very accustomed to using a single national ID. In many countries, however, promoting a standardized identity system is a real obstacle.”
Within this landscape, FinVolution’s international footprint spans Indonesia, the Philippines, and Pakistan, with brands licensed by local regulators: AdaKami under Indonesia’s Financial Services Authority and JuanHand under the Securities and Exchange Commission in the Philippines. Beyond lending facilitation, the company has rolled out Blu, a product that supports voice calls and text-based interactions across financial scenarios. Blu is now accessible in six countries across Asia and Latin America.
The Philippines offers a case study in localized adaptation. There, FinVolution received approval to establish a credit bureau regulated by the local Credit Information Corporation. By integrating data that reportedly cover about 70% of underserved consumers and deploying more than 50 tailored risk control models, the company launched an intelligent scoring system for nonbank customers, moving toward a more inclusive credit infrastructure.
As the company extends partnerships with digital banks, e-commerce platforms, and other traffic channels, its export of AI-driven risk tools and straight-through lending technology is likely to become its principal value proposition abroad.
Looking ahead, FinVolution aims to enter ten or more Asia Pacific markets by 2030 and to raise the share of international revenue to 50% by then, according to Bamboo Works. While demanding, these goals align with the company’s thesis that robust AI-driven systems, backed by talent and research communities, are prerequisites for scale.
A platform for talent and responsible innovation
The FinVolution Global Data Science Competition began as a way to test algorithms on hard, consequential problems. Ten years on, it functions as something broader, a platform that links academia and industry, cultivates young talent, and probes where cutting-edge research meets operational reality.
With deepfakes evolving rapidly and globalization accelerating, the through line remains the same: advance technology for good by building reliable defenses, sharing knowledge, and turning promising ideas into systems that safeguard users at scale.
As generative AI progresses, FinVolution’s task is twofold: continue strengthening safety by expanding from static image forensics into dynamic video and cross-scenario defenses, and apply these capabilities responsibly in new markets. The next decade of the competition, and of the company’s globalization, will be judged not only by technical breakthroughs but also by how well they translate into durable protections for consumers, institutions, and the financial systems that serve them.
This article was published in partnership with FinVolution Group.