Cryptic but Intelligent: How Homomorphic Encryption can Revolutionize Artificial Intelligence
Have you ever heard of homomorphic encryption? If not, I recommend you check it out because that relatively unknown security parading promises to become one of the most important developments to accelerate the mainstream adoption of artificial intelligence(AI) in the near future.
Data sharing is one of the biggest challenges with the adoption of AI solutions in the real world. While every week there are new AI models that pioneer breakthrough techniques, they hardly ever get applied against confidential datasets from regulated industries. The examples that we find in AI research literature of AI algorithms evaluated using public datasets such as the famous MNIST, they are rarely directly applicable in real world scenarios. That’s just a reality.
In this ultra-competitive, global economy, data has become a new form of currency and companies closely guard and protect their precious datasets frin anybody that can get extra competitive advantage from analyzing them. as a result, in most regulated industries such as financial markets, healthcare or defense, there are almost no public datasets that can be actively used by AI researchers in order to tackle mission-critical problems from that industry. Furthermore, most companies are reluctant to share their datasets with AI labs in universities because of fears that the information might be stolen. That bring us to homomorphic encryption.
What is Homomorphic Encryption?
The roots of homomorphic encryption date back to 2008, when an IBM researcher named Craig Gentry was able to perform a series of mathematical operations on encrypted data without needing to decrypt it. Gentry’s initial prototype tool the cyber-security world by storm and it is widely regarded as the first example of homomorphic encryption.
The biggest challenges with practical homomorphic encryption techniques is that they tend to require ridiculous amount of computing power. However, recent breakthroughs by companies such as Microsoft and IBM seem to be taking us closer to see homomorphic encryption techniques available in cloud platforms such as Azure or Bluemix.
Homomorphic Encryption and AI
Clearly, homomorphic encryption will allow organizations to share protected datasets so that they can be used in AI models. However, the benefits of homomorphic encryption for AI agents go beyond conceptual and into the core mathematical foundation of this new security technique. Today, we have practical homomorphic encryption techniques such as the Fan and Vercauteren Schemes that allow the execution of addition and multiplication operations on high degree polynomial ciphertexts in an algebraic ring. The thing is that achieving consistency in addition and multiplication helps to preserve the structure of the dataset. Considering that AI algorithms deeply care about the structure of the data, that type of homomorphic encryption technique ais key to allow the execution of AI models on encrypted datasets. Neural cryptography is another similar technique that is specifically optimized for machine learning algorithms using encrypted datasets.
homomorphic encryption has the potential of bridging the gap between enterprises and AI researchers. Using homormophic encryption techniques, companies can share confidential, encrypted datasets with partner companies so that they can be used in novel AI models. similarly, homomorphic encryption can help to enforce new levels of compliance and regulation in AI solutions without necessarily constraining progress.