Great to see that research in this classic field in image processing remains active. Most image processing research is based on deep learning methods. But it should be noted that the last section contains a comparison with a machine learning approach to de-noising (The performance of CNN with Median is the best among tested methods).
I think the term « device » is a bit misleading. I should have used client side instead. It doesn’t need to scan your messages between keyboard and encryption, the application handles your messages in plaintext and it can scan them without needing any decryption. The encryption is done just before sending the messages.
The EU has many types of legal acts and the European Commission is responsible for planning and proposing law. Also EU regulations are legal acts that apply automatically to all EU countries without needing to be transposed into national law.
For scanning encrypted messages, they can be scanned on device before encryption, like Apple’s approach to CSAM.
Imagine the cost of all the compliance procedures that a tech company should implement in the EU: GDPR, Digital Markets Act, and maybe in the future: Chat control. I think that the EU is putting barriers to entry that can backfire and harm its own tech companies, even if the initial plan is limiting the influence of US corporations for digital services.
The paper is mentioned also in the book "Skunk Works: A Personal Memoir of My Years of Lockheed". The Skunk Works' design team used the equations published by Soviet physicist and mathematician Petr Ufimtsev on electromagnetic wave reflection to create the plane.
Just to be fair also to some commenters, I think that the post had been edited after posting from what I remember ... so maybe the older comments are not very relevant.
> It involves bad practices and errors from multiple parties in a world that might seem
> foreign to the "Silicon Valley" world but paints an accurate picture of what
> development is for small IT companies around the world
Everybody makes mistakes even in the "Silicon Valley" world, but such problems cloud be easily caught by testing (which he did but it was restricted to the first page) and performing a simple dry-run.
I've integrated keycloak for SSO in our apps, we have a one-to-one mapping between keycloak realms and the tenants in our apps, works great and it can glue together many disparate IdP/auth solutions: onelogin, LDAP/AD, etc. At one moment we needed to implement a custom OTP method, so we developed a plugin by implementing a simple SPI (the auth flow is configurable).
It is a very stable piece of software, we have a combination of Terraform+ansible scripts to deploy it, and once it is done we forget about it until we need to upgrade the version using the same scripts. The biggest drawback IMHO is the documentation quality ...
Excuse my ignorance, I keep seeing those hefty fines on big tech companies, but does the EU targets also other types of industries ? For example Visa Mastercard duopoly, etc
Just to correct the info: the invasion of Iraq (a much weaker country) by the combined force of troops from the United States, the United Kingdom, Australia and Poland lasted just over one month … but it’s not really comparable situation here for the reasons you pointed.
I don’t think it’s a single or even a fixed collection of zero days, it’s an arms race that requires constant updates to the vulnerability catalog in order to be able to exploit the latest fully patched phones.
In discrete mathematics (combinatorics) there are definitely some tools and techniques but seemingly every problem is unique, I’m not quite sure that pattern matching is very useful in this subfield of mathematics ?
The problem with Apple’s approach to CSAM is that they use Neuralhash. Unlike other simple perceptual hashes, the failure modes and the collisions using this method are not well understood. I repeat here my previous comment in another thread : they use NN and triplet embedding loss, the exact same techniques used by neural networks for face recognition, so maybe the same shortcomings would apply here. For example a team of researchers found a 'Master Faces' that can bypass over 40% of Facial ID. Now suppose that you have such an image in your photo library, it would generate a ton of false positives and not just a single match with the NCMEC database.
Given that Apple technology uses NN and triplet embedding loss, the exact same techniques used by neural networks for face recognition, so maybe the same shortcomings would apply here. For example a team of researchers found a 'Master Faces' that can bypass over 40% of Facial ID.
Now suppose that you have such an image in your photo library, it would generate so many false positives …
I think that the used technology described in CSAM Detection Technical Summary (the so called NeuralHash) is a really bad idea, basically it uses the output of a neural network (likely a CNN) trained using the now classical triplet loss (see page 5).
The problem with those methods that they are not always reliable, for instance the same type of networks and same training procedure is used in neural face recognition that proved to be vulnerable (A recent example is the 'Master Faces' That Can Bypass Over 40% Of Facial ID, etc).
The analysts assume rationale behavior, but it should be noted that some factors such as money laundering distort the prices and
Toronto real estate market has a massive floods of opaque cash used for money laundering (at least $28.4 billion invested in Greater Toronto Area housing since 2008 according to reports).
The article mention : "Egypt fears its supplies will be reduced during the time it takes to fill the 74-billion-cubic-metre capacity reservoir." From information that I can find, the filling can take up to 7 years !
Unfortunately for this demo, they will successfully verify everyone. I was hoping for a real demo, in the past I had some interesting problems with selfie KYC checks because the photo in my passeport and my actual look are quite different …
Thank you, yes we are using this feature but I was thinking of something like creating separate isolated network for each group of machines. For example I manage two different cloud infrastructures that don’t share anything in common. It would be more convenient if I could place them in different tenants.
I use Tailscale every day I like the simplicity of setting up a new node. My problem is that I don’t know how to separate my nodes into isolated groups, so my question is does it support some form of multi-tenancy ?
I really like APL but I cannot imagine myself depending on a closed source tool-chain. I've heard that the GNU APL is order of magnitude slower than Dyalog APL, but I can't find a benchmark to confirm that.
Thank you for the link, yes we are aware of this particular issue.
For our engine we used only lossless mode and using this mode the decompressed binary image is identical to the original so the described problem cannot occur. Nonetheless this limit severely the usefulness of JBIG2.
We used it to compress scanned PDFs, using mixed raster techniques. Since the PDF natively supports JPEG2000 we used it for document background compression and we used JBIG2 for text image compression. The compression ratio was impressive but the document rendering was noticeably slower even with Adobe reader …
For WinMerge it should be noted that some web-sites out-there offer winmerge binaries for download and I bet that some of them are bundled with spyware (they are not affiliated with the official version, but even my tech-savvy friend was about to download winmerge the other day from a non official source, I noticed him that he should use winmerge.org).
Thank you for this info, as you mentioned CLIP is used for re-ranking DALL-E outputs, by itself it is just an (image, text) pairs classification network.
It's a recurring theme in OpenAI research, they become more and more closed.
For instance their latest model called DALL·E hit the headlines before the release of the paper. Needless to say, the model is not available and no online demo has been published so far.
To benefit from the acceleration, one should use the Apple implementation of BLAS/LAPACK: the Accelerate framework, for the moment I think numerical python libs link against OpenBLAS ? But I'm not 100% sure.
After reading the comments and the blog post, I think that the requirements boils down to fast persistence to disk, minimum dependencies and fast test-runs. Fortunately the data is very small 150MB and it fits very easily in memory.
According to the post the data changes often so they need to write the data many times in a second. But I'm not sure why do they need to flush every time the entire 150MB ? Why not structure the files/indexes such that we write only the modified data ?