Shulou(Shulou.com)06/03 Report--

As development teams scramble to develop AI tools, training algorithms on edge devices is becoming more and more common. Joint learning (Federated learning), a subset of distributed machine learning, is a relatively new approach that allows companies to improve their AI tools without explicitly accessing raw user data.

Joint learning, conceived by Google in 2017, is a decentralized learning model through which algorithms can be trained on edge devices. With regard to Google's "machine learning on device (on-device machine learning)" method, the search giant pushed its predictive text algorithm to Android devices, aggregated the data, and sent a summary of the new knowledge back to the central server. In order to protect the integrity of user data, this data is transmitted through homomorphic encryption or differential privacy, which is a way to add noise to the data to blur the results.

In general, through joint learning, the AI algorithm can be trained without identifying any specific data of individual users. In fact, the original data will never leave the device itself, and only summarized model updates will be sent back. These model updates are then decrypted after being delivered to the central server. The test version of the updated model is then sent back to the selected device. After repeating this process thousands of times, the AI algorithm is significantly improved without compromising user privacy.

This technology is expected to make waves in the medical field. Medical startup Owkin, for example, is currently exploring joint learning. To take advantage of patient data from multiple medical institutions, Owkin uses joint learning and uses data from different hospitals to build the c algorithm. This can have far-reaching implications, especially because it is valuable for hospitals to share disease progression data with each other while maintaining the integrity of patient data and complying with HIPAA regulations. Healthcare is by no means the only sector adopting this technology; joint learning will increasingly be used by self-driving car companies, smart cities, drones and fintech organizations. Several other joint learning startups are about to go public, including Snips, S20.ai and Xnor.ai, which was recently acquired by Apple.

Potential problem

Man in the middle attack (Man-In-The-Middle Attacks)

Given that these AI algorithms are worth a lot of investment, these models are expected to be profitable targets for hackers. Evil hackers may try to carry out man-in-the-middle attacks. However, as mentioned earlier, companies can make it difficult for hackers to do this by adding noise and aggregating data from various devices, and then encrypting the aggregated data.

Model poisoning (Model Poisoning)

Perhaps more worrying are the attacks that poisoned the model itself. It is conceivable that hackers can destroy the model through their own devices or by taking over the devices of other users on the network. Ironically, hackers entering through the back door are masked to some extent because joint learning gathers data from different devices and sends encrypted summaries back to the central server. Therefore, it is difficult, if not impossible, to identify the location of the exception.

Bandwidth and processing limitations

Although machine learning on the device effectively trains the algorithm without exposing the original user data, it does require a lot of local power and memory. The company tries to circumvent this problem by training its AI algorithm at the edge only when the device is idle, recharged or connected to the Wi-Fi; however, this is an eternal challenge.

The influence of 5G

With the global expansion of 5G, edge devices will no longer be limited by bandwidth and processing speed. According to a recent report by Nokia, 4G base stations can support 100000 devices per square kilometer. The upcoming 5G base station will support up to 1 million devices in the same area. With enhanced mobile broadband and low latency, 5G will provide energy efficiency while promoting device-to-device communication (D2D). In fact, it is predicted that 5G will bring a 10-100-fold increase in bandwidth and a 5-10-fold reduction in latency.

As 5G becomes more popular, we will experience faster networks, more endpoints, and larger attack surfaces, which may attract an influx of DDoS attacks. 5G also has slicing function, which can easily create, modify and delete slices (virtual network) according to the needs of users. It remains to be seen whether such network slicing components will allay security concerns or lead to a series of new problems, according to a study of the destructive power of 5G.

All in all, new concerns have emerged from a privacy and security perspective; however, the fact remains that 5G is ultimately a boon for joint learning.

Https://ai.51cto.com/art/202006/619339.htm

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.