Masking is a workhorse privacy technique: add randomness to a value so an observer cannot read it, then strip the randomness later. Inside fully homomorphic encryption it gets used to protect intermediate values during a computation. But there is a subtlety — sometimes the later computation still needs to know the sign of the masked value, and naive masking destroys it.
US12206757B2, “Efficient random masking of values while maintaining their sign under fully homomorphic encryption (FHE),” granted to IBM on January 21, 2025, is a patent about that precise tension. Classified under H04L 9/008, it claims masking that randomizes a value for privacy while preserving whether it is positive or negative.
This is exactly the kind of fine-grained, almost surgical claim that characterizes a maturing field. Nobody files a patent on sign-preserving masking unless they are deep enough into building real FHE computations to have hit the problem in practice. Comparisons, thresholds, and many machine-learning operations branch on sign; if masking obliterates the sign, those operations break or leak.
The word “efficient” in the title is also load-bearing, as it tends to be in FHE patents. Anything can be done expensively under homomorphic encryption; the engineering value is in doing it cheaply enough to be usable inside a larger encrypted computation. An efficient sign-preserving mask is a small primitive that bigger private algorithms can call without paying a prohibitive cost each time.
Per the desk's discipline: issued grant (B2), not an application; a method claim, not a product. IBM's long-running FHE research and its HElayers/HELib tooling are the obvious context, and the inventor team (Adir, Masalha, Aharoni) recurs across IBM's homomorphic filings.
For the reader gauging where FHE stands, the granularity of the claim is the signal. When a field's patents shift from “how to compute on ciphertext at all” to “how to mask a value while keeping its sign efficiently,” it has moved from feasibility to optimization — the unglamorous, detailed phase where a technology actually gets engineered into something deployable.