On 10/12/25 15:23, macchina wrote:
> ma poi chatcontrol, ma come la vorrebbero implementare a livello
> tecnico? mettendo degli spyware in ogni dispositivo?
usando algoritmi lato client
ci sono vari sistemi proposti, quello che va per la maggiore credo sia
applicare un modello col machine learning sulle immagini e video
presenti sul telefono, ora tutto questo si chiama AI e così la gente si
confonde, non capisce più cos'è e va in para.
Ero già pronto a vedere cosa succedeva dopo il voto, solo che la
Germania ha messo il veto e quindi ChatControl si è bloccata di nuovo..
fine della seconda stagione.
vediamo cosa succedera' prossimamente
cmq, qui una piccola spiegazione (in inglese) e qualche paper che spiega
in modo molto generico il "come far funzionare" questa cosa
Perceptual hashing
> Perceptual hashing. Hashes are specialized algorithms capable of digesting a large
> input file and producing a short unique “fingerprint” or hash. Many scanning systems
> make use of perceptual hash functions, which have several features that make them
> ideal for identifying pictures. Most importantly, they are resilient to small changes
> in the image content, such as re-encoding or changing the size of an image. Some
> functions are even resilient to image cropping and rotation.
> Perceptual hashes can be computed on user content and then compared to a
> database of targeted media fingerprints, in order to recognize files that are identical
> or very similar to known images. The advantage of this approach is twofold: (1)
> comparing short fingerprints is more efficient than comparing entire images, and (2)
> by storing a list of targeted fingerprints, providers do not need to store and possess
> the images.
Machine Learning
> Machine Learning. The alternative approach to image classification uses machine-
> learning techniques to identify targeted content. This is currently the best way to
> filter video, and usually the best way to filter text. The provider first trains a machine-
> learning model with image sets containing both innocuous and target content. This
> model is then used to scan pictures uploaded by users. Unlike perceptual hashing,
> which detects only photos that are similar to known target photos, machine-learning
> models can detect completely new images of the type on which they were trained.
> One well-known example is the face detector used in iPhones to detect faces on which
> to focus the camera
> While these two scanning technologies operate differently, they share some com-
> mon properties. Both require access to unencrypted content for matching. Both can
> detect files that the system has not seen before, though perceptual hashing is limited
> to detecting files that differ only slightly from images it has seen before. Both
> methods have a non-zero false positive rate. Both methods also rely on a propri-
> etary tool developed from a corpus of targeted content, which may be controlled by a
> third party. Some scanning techniques also use proprietary algorithms (for example,
> Microsoft’s PhotoDNA is available only under a nondisclosure agreement). Finally,
> regardless of the underlying technology, either method can be treated as a black box
> that inputs an unencrypted image and outputs a determination of whether it is likely
> to contain targeted material.
> These commonalities result in scanning based on both methods having very similar
> security properties. Both methods can be evaded by knowledgeable adversaries
> and both methods can be subverted in similar ways
going deep:
# paperz
Bugs in our Pockets: The Risks of Client-Side Scanning
https://www.cs.columbia.edu/~smb/papers/bugs21.pdf
Client-Side Scanning
What It Is and Why It Threatens Trustworthy, Private Communications
https://www.internetsociety.org/wp-content/uploads/2020/03/2022-Client-Side-Scanning-Factsheet-EN.pdf
Leakage-Abuse Attacks Against Searchable Encryption
https://eprint.iacr.org/2016/718.pdf
# linkz
https://www.eff.org/deeplinks/2019/11/why-adding-client-side-scanning-breaks-end-end-encryption
