Good article - one takeaway is that any redaction process which follows a fixed algorithmic sequence (convolutions, transformation filters, etc) is potentially vulnerable to a dictionary attack.
I see what you mean, but FWIW “fixed” doesn’t sufficiently constrain or describe it. For example, filling a rectangle with black or random pixels is a fixed algorithmic sequence, same might go for in-painting from the background. The redaction output simply should not be a function of the sensitive region’s pixels. The information should be replaced, not modified.
If we're talking "proper math terms", if it "returns a random number" it isn't a function. In math, the value of a function can't change unless the arguments change. If you evaluate it repeatedly with the same argument(s) you'll always get the same result.
Yes, you’re right, good point. Maybe there is no one good term for this case (but given the ocean of terminology, I’d be slightly surprised). ‘Not a function’ also isn’t the right term here because functions of x that returns a constant are okay - they just don’t depend on x. Hashed random functions are true functions but are designed to be non-invertible, so maybe non-invertible (or irreversible) is a decent single term for what @vunderba meant. Other terms that broach it might be ‘non-injective’ and ‘entropy-reducing transform’. I suspect those aren’t technically strong enough for the kind of information loss we need in this context.
