CEO of NTT Research, leading research in physics & informatics, cryptography & information security, and medical & health informatics.
Bad news in cybersecurity gets a lot of attention. Headlines about data breaches and new forms of malware tend to outweigh any good news that comes from the field of cryptography.
Because this field is highly mathematical and operates on a lengthy time horizon, progress can be difficult to convey. As I discussed previously, it took about 15 years for the capabilities of attribute-based encryption (ABE) to move from a technical paper into the standards world and now possible deployments.
However, academic research is a leading indicator. Recent developments include a solid theoretical foundation for software obfuscation, a new defense for public-key cryptosystems and several ways to make machine learning (ML) more trustworthy.
Obfuscation Finds Terra Firma
For 20 years, cryptographers have been working on ways to make software “unintelligible” while preserving its functionality. The technical foundation for that black-box capability is called “indistinguishability obfuscation” (iO). It’s a powerful concept that has appeared in more than 100 papers.
Despite the potential applications, however, the constructions of iO itself were on shaky ground. Dr. Amit Sahai, the director of the Center for Encrypted Functionalities at UCLA, is the senior author of a breakthrough paper published in August 2020 titled, “Indistinguishability Obfuscation from Well-Founded Assumptions,” which was co-authored by cryptographers associated with NTT Research and the University of Washington. In a statement regarding the paper, Sahai said: “For years, the mathematical foundations of indistinguishability obfuscation were, quite frankly, a mess.”
The authors admit that it will take a lot more work to make iO efficient and practical. However, this mathematical proof should settle doubts about whether iO really works or even exists. It also means that going forward, we will be able to take advantage of many strong functionalities that could not be realized without iO.
Defense Against Active Attackers
In another paper that won one of the three Best Paper awards at August’s Crypto 2020, which is organized by the International Association for Cryptologic Research (IACR), cryptographers demonstrated a powerful enhancement to the security of widely used public-key cryptosystems. The paper, titled “Chosen Ciphertext Security from Injective Trapdoor Functions,” provides a way to transform a trapdoor function — which is the most basic and core functionality of current public-key cryptosystems — into one with the strongest security that protects against active or interactive attackers.
“Chosen ciphertext security is so critical that only cryptosystems meeting these standards should be considered for deployment today,” noted Dr. Brent Waters, NTT Research scientist and one of the paper’s co-authors.
How Trustworthy Is Machine Learning?
Cryptography is also relevant to ML. In a recent keynote address at our virtual summit, Dr. Shafi Goldwasser, the director of the Simons Institute for the Theory of Computing at the University of California, Berkeley, summarized recent research indicating three ways that cryptography can strengthen ML.
In genome-wide association studies (GWAS), for instance, it is desirable to separate genomic samples from patient identity. Goldwasser summarized the results of tests involving two possible approaches — multiparty computation and homomorphic encryption. Both work, but contrary to the common expectation, homomorphic encryption — which enables computation on data that remain encrypted — proved to be the faster of the two.
Another challenge is robustness. An ML algorithm should, in fact, be able to distinguish a cat from a dog or a stop sign from a yield sign, but an “adversary” could prevent it from doing so. Goldwasser summed up another paper showing that the right defense against such misclassification first allows the learner to abstain from deciding and then requires that an algorithm have a low rejection probability on P (training data distribution) and a low error rate on Q (test data distribution).
ML also needs to be verifiable. A model may be telling you whether or not to grant a loan or issue bail, but did the model actually do the work or just flip a coin? According to a third paper co-authored by Goldwasser, you can indeed verify an ML model quickly and, more importantly, without access to the base data.
These examples create a counter-narrative to the depressing news of cybersecurity breaches. Thanks to the breakthrough work on iO, software applications could now become bona fide “black boxes,” fully functional yet impenetrable to attackers. The possibility of giving public-key infrastructure (PKI), which underpins web encryption and internet data privacy, the strongest possible protection is also a big deal. As for ML models, it is encouraging to hear of efficient approaches to privacy, robustness and verification.
Basic research is not product development, and sometimes years pass before deployments, but it is good to see even in this limited sample how the foes of cybersecurity face serious cryptologic obstacles that are becoming ever-increasingly difficult to overcome.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?