mirker

20 ms·

181.
by mirker 5y ago
Yes, a great current example is consumer motherboards with Intel’s I225V network controller. The controller is unable to run at 2.5Gbps without a fix. The controller was fixed in revision 3, but a huge number of boards are shipping with rev
182.
by mirker 5y ago
It’s the same thing but more time consuming. If you knew this was happening, you’d create N accounts, match with the same person, and average queries over account pairs. Also, if the error is constrained to be the same, then the attack is r
183.
by mirker 5y ago
It’s only a partial solution; you still need to limit the queries. Otherwise, just sample a large number of coordinates and average out the noise.
184.
by mirker 5y ago
The notion of fairness you are discussing is 1) vague 2) very similar to standard training. If you have an optimization problem that represents your fairness, it’d be more useful to just compare that with the other problems in the literatur
185.
by mirker 5y ago
Ok, a lot to unpack here. The viewpoint I am presenting is not novel and is what the paper is about. These “academic” notions of fairness are things you can prove for models, unlike ad-hoc variants that “feel” fair. So it’s not “my” viewpoi
186.
by mirker 5y ago
It depends on the definition of fairness, which is what I’m trying to argue for. Practically speaking, heavy correlation and casual relationships are similarly problematic. If you correlate enough of the variables, you can limit your pool t
187.
by mirker 5y ago
That’s not really true. Say you want to know if you should approve someone for insurance. SELECT AVG(profit) WHERE feature == $user_feature; Say “feature” is education level and then suppose education level is correlated with race or gender
188.
by mirker 5y ago
Have we even figured out what bias we care about? Race, gender, age, etc. are some potential problems, but is that it? These sorts of problems are often formulated theoretically (“suppose we want fairness with respect to variable Z”). It se
189.
by mirker 5y ago
Most will not pass up an opportunity to keep their job or get promoted, so I think the practical answer lies elsewhere in the research pipeline.
190.
by mirker 5y ago
The way I think about it is that the attacker has plenty of time (years) to either guess or discover the viability of an attack. For example, inspecting packets and reverse engineering, correlating photos with police visits, creating a hash
191.
by mirker 5y ago
The privacy aspect does seem novel, and that paper evaluates the various aspects of privacy and the quality of perceptual hashes. However, anything concerning privacy (of client or server data) is orthogonal to the vision methods underpinni
192.
by mirker 5y ago
> Apple genuinely believe this is a novel and unique method to catch CSAM without invading people's privacy What’s novel about this? The technique and issues with it are fairly obvious to anyone with experience in computer vision. I
193.
by mirker 5y ago
Citations follow some power-law type of distribution, so the 10% (or whatever) of useful ones are basically the high impact papers, anyway. It would be more surprising if citations and usefulness were anti-correlated in some way. Back to th
194.
by mirker 5y ago
For what it’s worth, an attack similar to above is now demonstrated. https://news.ycombinator.com/item?id=28219068
195.
by mirker 5y ago
For transferring, I mean specifically that it is known that attacks transfer in a general sense e.g., https://www.usenix.org/system/files/sec19-demontis.pdf It’s not easy in this setting due to the information asy
196.
by mirker 5y ago
I was under the impression that the hashes were the only thing transferred, as the source image is encrypted. (Edit: On second thought, manual reviewing needs the source image.) Anyhow, adversarial attacks transfer reasonably well across mo
197.
by mirker 5y ago
I would guess that the main point has to be hardware accelerated features, such as ray-tracing. I agree though that it seems pointless to buy a budget GPU when it’s basically a scaled up iGPU. Perhaps it makes sense if you want a mid-range
198.
by mirker 5y ago
My favorite example is compound interest. Imagine having to solve a compound interest test to get a credit card. It seems like a basic precaution to me (like a driving license), but yet I’m not convinced all in the US can do it.
199.
by mirker 5y ago
It’s an attack vector because, either directly or indirectly, people will get access to the hash function. Therefore, they will have a testbed to launch attacks against the hash. Once they learn an attack, they will hash your harmless pictu
200.
by mirker 5y ago
Convolutional Neural Network e.g., deep learning
201.
by mirker 5y ago
SQL expresses programs via the relational model. Pandas is similar. I’m not an expert on the areas, but it seems that relational modeling is the best tradeoff between performance and abstraction. To your point, people did have a “programmin
202.
by mirker 5y ago
But you cannot know that a-priori so it’s either an attack vector for image manipulation or straight up false positives. Assume we had this perfect hash knowledge. I’d create a compression algorithm to uniquely map between images and the 25
203.
by mirker 5y ago
If Apple is indeed using CNNs, then I don’t see why any of the black-box adversarial attacks used today in ML wouldn’t work. It seems way easier than attacking file hashes, since there are many images in the image space that are viable (e.g
204.
by mirker 5y ago
If images have cardinality N and hashes M and N > M, then yes, by pigeonhole principle you will have collisions regardless of hash function, f: N -> M. N is usually much bigger than M, since you have the combinatorial pixel explosion.
205.
by mirker 5y ago
You need to provide an example of a highly cited paper with low clarity otherwise the clarity concept is vague. Even if it is true that papers are unclear, the incentives for academics is to get prestige, which follows from getting papers a
206.
by mirker 5y ago
It’s impossible to argue meaningfully without an example (for a field with tens of thousands of papers a year), but the purpose of academic papers is to convince experts of the ideas. If the paper is clear enough for an expert, it is servin
207.
by mirker 5y ago
Thanks for the post. One question: One of the advantages about having a clean design is that performance is easier to optimize, since the 80%/20% rule of performance becomes much more obvious. How true was this in your experience? Were
208.
by mirker 5y ago
Latex is my personal example of bad compiler messages, as well. The other day I added a duplicate bibtex entry and/or added two keys for the same entry. The error messages are a dump of the entire bibliography rather than a simple li
209.
by mirker 5y ago
I’d be interesting to study how people read Latex compiler errors, which I find to be unhelpful and an extreme case of bad error messages. I think part of the problem is an extensive use of macros, which expand out into gibberish with littl
210.
by mirker 5y ago
How much of that hardness is necessary, though? Backward compatibility is quite costly.

More ›