A multinational company discovered its confidential system data had been sold on the dark web. However, a review of system access logs revealed no unusual download activity. Primidea was engaged to investigate.

Recognizing the system’s open access and the lack of abnormal download logs, our team hypothesized that the data had been stolen by photographing the screen. Using three months of office surveillance footage, we employed object detection and pose estimation techniques to create training samples. By integrating these with open-source posture datasets, we trained an action classification model. This model allowed us to accurately analyze employee actions in the footage, ultimately identifying the individual who had photographed the confidential system pages.

Primidea’s solution enabled the client to quickly trace the source of the data breach, saving significant time and resources while preventing further leaks.

The audit team of a multinational company suspected distributors of submitting duplicate in-store display photos to claim multiple display and promotion fees. However, with thousands of SKUs and millions of photos uploaded monthly, manual detection was impractical.

Primidea found that the challenge lay not in object recognition but in identifying similar scenes. Elements like promotional posters, packaging, and store signage—which occupied large proportion of the photos—posed significant interference to accurate detection.

Our technical team innovatively combined semantic segmentation to eliminate these distractions, focusing instead on comparing store or interior layouts. The solution efficiently identified duplicate scene photos, greatly enhancing the client’s audit efficiency and streamlining their management processes.

A U.S.-based insurance company faced an uptick in full policy surrender requests. Investigations revealed rogue agents were falsely claiming affiliations with regulators or special resources to mislead customers into claiming full refund. These agents charged high service charges or directed customers to purchase competitor products, even filing fraudulent complaints to pressure the company into refunding premiums.

Primidea analyzed thousands of phone recordings provided by the client. We cannot distinguish between customer and malicious agents by analyzing the phone numbers alone. Using voiceprint recognition, however, we managed to extract and compare voice features to identify matches across different recordings. This process successfully flagged hundreds of calls likely made by the same rogue agents.

The findings provided the client with strong evidence to pursue claims and protect their business interests.

A financial institution identified that certain sales personnel had exploited system vulnerabilities to access confidential customer information. With access logs for some periods no longer available, the company opted to review Virtual Desktop Infrastructure (VDI) recordings from over 5,000 employees over the past one year to identify suspicious activity and quantify the scope of the data breach.

Primidea deployed advanced machine vision algorithms to tackle the vast dataset efficiently. By analyzing system interface features and combining template matching with object detection, we accurately pinpointed relevant screen content. Optical Character Recognition (OCR) technology was then employed to extract critical data fields from the identified frames.

This approach successfully revealed dozens of employees involved in unauthorized access and uncovered more than 10,000 compromised customer records. Our solution provided the client with comprehensive insights to strengthen their data security framework and enhance compliance protocols.

A whistleblower alerted a multinational company to allegations that over ten employees were involved in leaking and selling confidential company data. The company launched an internal investigation and engaged Primidea to analyze the employees’ electronic data.

Upon examining the work computers of the implicated employees, Primidea uncovered hundreds of thousands, even millions, of cached images on each device. A manual review of this scale would have been highly inefficient and risked overlooking critical evidence. To streamline the process, Primidea’s technical team implemented an AI-powered image classification system.

The team first categorized and prioritized images most likely to contain relevant evidence, identified key categories, and created a robust training dataset. Leveraging this dataset, we developed a precise image classification model. With AI, investigators were able to review the immense volume of images in just a few days.

The investigation uncovered pivotal evidence, including:

• Screenshots of employees sharing company data via WeChat.
• Screenshots of employees receiving bribes through WeChat or bank transfers.

These findings were instrumental in helping the company address the breaches and mitigate further losses, demonstrating the power of AI in streamlining complex investigations.