Data privacy and audit compliance of supply chain finance warehouse receipt pledge based on WSN federated edge learning

In the world of modern finance, small and medium-sized businesses often struggle to get the loans they need to grow. One common solution is warehouse receipt pledge financing, where a company uses its stored inventory—like grain, metal, or manufactured goods—as collateral to secure a loan from a bank. For this system to work, the bank must trust that the goods actually exist, that they are in good condition, and that they haven't been swapped or tampered with. Traditionally, verifying this relies on manual inspections and periodic inventories, which are slow, prone to error, and leave gaps in knowledge. Furthermore, the data describing these goods and the financial details of the loans are highly sensitive. Banks and warehouses often treat this information as a secret, creating isolated pockets of data that cannot be easily shared to build a better picture of risk. This creates a dilemma: how can banks collaborate to detect fraud and manage risk without forcing companies to reveal their private business secrets?

To solve this, researchers from Shandong University of Technology have developed a new system that combines physical sensors, a method for training artificial intelligence without sharing raw data, and a digital ledger for verification. They focused on a scenario where multiple warehouses, each holding different types of goods for different companies, need to work together to spot risky pledges. The team built a system that uses wireless sensors to watch the goods, a smart learning method that keeps data local, and a secure way to prove the rules were followed without exposing the secrets. Their work demonstrates that it is possible to create a highly accurate risk model that respects privacy and satisfies strict regulatory audits, all while operating on the limited power of small sensor networks.

The researchers began by addressing the physical reality of the pledged goods. They equipped warehouses with a network of small, wireless sensors that constantly monitor temperature, humidity, weight, and the location of the items. These sensors act like a continuous digital eye, capturing the real-time state of the inventory. Instead of sending this raw, sensitive data to a central server where it could be stolen or leaked, the system processes the information locally at the warehouse. It then creates a "digital warehouse receipt," a secure record that links the physical condition of the goods to the business details of the loan. This receipt is stored on a blockchain, a type of digital ledger that is extremely difficult to alter, ensuring that the record of what is in the warehouse matches the physical reality. If the sensors detect a sudden drop in temperature or a change in weight that doesn't match the records, the system immediately flags an anomaly, alerting both the warehouse operator and the bank.

Once the physical data is secured and converted into trusted digital records, the researchers tackled the challenge of collaborative risk modeling. In a traditional setup, banks would have to share their private data to train a computer model that predicts which loans are risky. In this new approach, the data never leaves the warehouse. Instead, the researchers used a technique called federated edge learning. Imagine a group of students in different classrooms who are all trying to solve the same difficult math problem. Instead of sending their notebooks to a central teacher to be graded, each student works on their own problem, figures out the general rule they discovered, and sends only that rule to the teacher. The teacher combines all the rules to create a master guide and sends it back. The students then use this improved guide to refine their own work. In this system, each warehouse acts as a classroom, training a local version of the risk model on its own data. They only share the "rules" they learned—the mathematical updates to the model—rather than the raw data itself.

To make this process efficient and reliable, the researchers added several layers of intelligence. They designed a selection process that chooses which warehouses participate in each round of learning based on their trustworthiness and their remaining battery power, ensuring that only reliable, well-powered nodes contribute. They also introduced a mechanism to keep the local models from drifting too far apart, which is a common problem when different warehouses have very different types of goods. To further protect privacy, the researchers wrapped the model updates in two layers of security. First, they added a small amount of mathematical noise to the updates, making it impossible to reverse-engineer the original data. Second, they encrypted the updates so they could be combined while still locked in a secure format. Only a trusted authority could unlock the final combined result. This ensures that even if someone intercepts the data being sent between warehouses, they cannot see the underlying business secrets.

Finally, the system includes a built-in audit trail that satisfies regulators. Every step of the learning process is recorded on the blockchain, creating a transparent history that shows who participated, what rules were followed, and how much privacy protection was applied. Smart contracts, which are self-executing agreements on the blockchain, automatically check if the process complied with all regulations. If a warehouse attempts to deviate from the protocol or if the privacy limits are exceeded, the system can detect it and alert the authorities. This creates a closed loop where privacy, accuracy, and compliance are all monitored in real time.

The results of the study show that this approach is highly effective. In their tests, the new method achieved an accuracy rate of 94.8% in identifying risks, which was significantly higher than existing standard methods. Even when the data from different warehouses was very different from one another—a condition that usually confuses computer models—the new system remained robust, maintaining high accuracy where other methods failed. The researchers also found that the system could operate under strict privacy conditions. When they increased the level of privacy protection to its maximum, the system still maintained an accuracy of 83.6%, whereas other methods dropped to 72.5%. While the security measures did require more computing time and energy, the trade-off was minimal in terms of the final result. The study concludes that by integrating wireless sensors, secure learning, and blockchain verification, it is possible to build a supply chain finance system that is not only smarter and more accurate but also respects the privacy of every participant and meets the rigorous demands of financial regulators.