No answer yet from what I’ve seen. Possibly Mess up from an unreviewed vibe coded mess. Or worse yet , intentional.

you are viewing a single comment's thread
view the rest of the comments
[–] 244 points 1 day ago* (42 children)

https://developers.google.com/ml-kit

ML Kit’s processing happens on-device. This makes it fast and unlocks real-time use cases like processing of camera input. It also works while offline and can be used for processing images and text that need to remain on the device.

https://developers.google.com/ml-kit/vision/doc-scanner

The entire document scanner flow operates on-device.

It's on-device scanning and OCR and stuff. I fail to see how this is a privacy concern? Your documents aren't leaving your device and they're not being uploaded to Google servers. This really seems like a false moral outrage imo. Y'all really wanna hate Proton because the CEO is a dickbag but c'mon man, this doesn't scream vibe coded, low quality, or insecure/non-private to me. I'm guessing this is just to allow you to upload documents to Proton using your phone as a scanner.

  • source
  • hideshow 42 child comments
  • [–] 94 points 23 hours ago (2 children)

    Concerning

    https://www.privacyaffairs.com/proton-drive-google-ml-kit/

    ML Kit isn’t silent, though. Google’s data disclosure for ML Kit lists what its Android libraries collect for diagnostics and usage analytics:

    Device: maker, model, OS version and available ML hardware.

    App: package name and version.

    Identifiers: a device identifier for features delivered through Play services, like this scanner, plus per-installation identifiers.

    Usage: performance figures, settings, input and output sizes, events such as start-ups, model downloads and detections, and error codes.

  • source
  • parent
  • hideshow 2 child comments
  • [–] 30 points 18 hours ago* (1 child)

    Kudos for actually looking into this at all and not just having feelings on the subject.

    Having said that, that usage data is not sold or made available for non-development purposes: https://developers.google.com/ml-kit/android-data-disclosure

    Overview of data encryption and data sharing for ML Kit Android SDKs:

    Data encryption For the collected data listed on this page, ML Kit encrypts the data in transit using HTTPS.
    Data sharing For the collected data listed on this page, ML Kit does not transfer this data to third-parties.

    I also haven't proven this out, but I think the devs should be able to block that behavior in the apps manifest and prevent the telemetry uploader from running.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 23 points 11 hours ago* (last edited 11 hours ago)

    Having said that, that usage data is not sold or made available for non-development purposes:

    Not yet. The usual pattern of Google is to ship something good for free and once it has a sufficient spread, extract the shit out of it, Human Rights be damned.

  • source
  • parent
  • [–] 91 points 1 day ago* (4 children)

    Thank you for the additional context!

    Proton has issues as a company but it feels like 50% of the negative headlines or coverage they get ends up not being an accurate representation.

    The other 50% sucks tho, which is a bummer

  • source
  • parent
  • hideshow 4 child comments
  • [–] 19 points 19 hours ago* (last edited 19 hours ago) (3 children)

    It really is. A spotty record like this is simply not good enough for a company putting forth privacy and security as its chief features and product differentiators. You're either all in on those things or you're not.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 4 points 11 hours ago (2 children)

    I know the issues with the man at the top, but what spotty record in regards to the actual product are there?

  • source
  • parent
  • hideshow 2 child comments
  • [–] 0 points 8 hours ago (1 child)

    My personal concerns include their longstanding neglect (years) of missing features, lack of Linux support on some of their offerings, concerning ToS changes, a stated philosophy of expediency over security and other goals, and appalling CEO conduct.

  • source
  • parent
  • hideshow 1 child comment
  • [–] [S] 36 points 1 day ago (6 children)

    Is there a guarantee ML kit stays on device or we just taking Googles word for it ?

  • source
  • parent
  • hideshow 6 child comments
  • [–] 18 points 1 day ago (3 children)

    You can sniff packets heading out of the device if you're so invested. Proton should frankly be the ones testing and making that statement clear, but you can too.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 3 points 19 hours ago (2 children)

    MLK phones home, so it's more a matter of taking Google's word that they're not sending your data back.

  • source
  • parent
  • hideshow 2 child comments
  • [–] 32 points 1 day ago (1 child)

    It would be pretty easy to demonstrate the opposite, a security researcher could do it in a jiffy, so it wouldn't make sense for Google to lie about it.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 19 points 22 hours ago* (last edited 22 hours ago)

    Yes, but also no, never know what G will do next, once you include their libraries terms can change on a whim.

    ETA: Apparently terms already permit exfiltration of identifiers and 'performance figures, settings, input and output sizes, events such as start-ups, model downloads and detections, and error codes.' (thanks BrilliantBadger) So simple detection won't work.

  • source
  • parent
  • [–] 13 points 1 day ago (13 children)

    It's not against proton.. But can you trust anything that is "on device" from G?

    Go offline, scan, encrypt, zip with password, purge the original file somehow then go online.. Even then you can't be sure there ia not some thumbnail, copy, screenshot being made and send to G

  • source
  • parent
  • hideshow 13 child comments
  • [–] 12 points 1 day ago (12 children)

    It's trivially easy to prove it works offline, and it would be also very easy to sniff traffic on device or MITM yourself to verify traffic isn't flowing to Google.

    As a closed-source package it's not possible to inspect the source but it would be very easy to catch them in the act, even with it as a black box.

  • source
  • parent
  • hideshow 12 child comments
  • [–] 3 points 19 hours ago (11 children)

    MLK phones home, so it's more a matter of taking Google's word that they're not sending your data back.

  • source
  • parent
  • hideshow 11 child comments
  • [–] 1 point 18 hours ago (10 children)

    Not really, you can tcpdump it or MITM yourself and read the data that's being sent.

    The app devs can also block the telemetry uploaded from running, which should be inspectacle from apps manifest.

  • source
  • parent
  • hideshow 10 child comments
  • [–] 0 points 18 hours ago (9 children)

    Not an expert, but can you if it's encrypted?

    The app devs can also block the telemetry uploaded from running

    Yeah but are they? Especially considering the recent turns proton has taken.

  • source
  • parent
  • hideshow 9 child comments
  • [–] 2 points 18 hours ago (8 children)

    Yeah you have to get the TLS cert so it's annoying but you can definitely decrypt HTTPS traffic via tcpdump. It might be extra annoying over Android but it's definitely possible somehow.

    Worst case you fake the cert and MITM your device and sniff as you pass it upstream. It's a lot of work but...

  • source
  • parent
  • hideshow 8 child comments
  • [–] 1 point 11 hours ago (3 children)

    You arent getting google's private cert. You can decrypt data they send. Not the other way around. HTTPS encryption is asymmetric. You cannot decrypt data you yourself encrypt that easily. Their TLS cert provides the encryption, not decryption.

    That being said, blocking the HTTPS data collection endpoint is a lot more trivial to do.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 2 points 8 hours ago* (2 children)

    You don't need Google's private cert to decrypt HTTPS traffic, you just need the keys from the TLS handshake your client performs. As you are the client, you can sniff your own traffic. I've personally done this using wireshark. I've already provided one link from a trustworthy source, feel free to search for more if you choose.

    https://support.atlassian.com/atlassian-knowledge-base/kb/how-to-capture-http-traffic-using-wireshark-fiddler-or-tcpdump/

  • source
  • parent
  • hideshow 2 child comments
  • [–] 1 point 4 hours ago (1 child)

    Does this not state you need the actual clienr application to be involved in the process? Where are you getting the session key here? Its not a browser.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 1 point 3 hours ago

    You just need access to the client itself, which in this case if you cared enough to go this far, could be a sample app that you're just using to test MLKit's behavior. Custom app means you can do whatever you want as part of the TLS handshake, including dumping the negotiated keys to a logfile. Again, it would be non-trivial work, but it's absolutely possible. If anyone was sufficiently motivated, they could do it.

  • source
  • parent
  • [–] 1 point 18 hours ago (3 children)

    I don't get it. The whole point of encryption is that only the recipient can decrypt it, regardless of what role you play in the middle. Unless you're reading it before the encryption happens.

    Not saying it's bullshit, just not jiving with my very basic understanding of this.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 2 points 18 hours ago* (last edited 18 hours ago)

    For MITM attacks you need to own the network you're on and then you have the client connect to you with a cert the client will trust (by adding to the client's trusted cert bundles or or or) and then sniff as you proxy upstream to the real intended host.

    This is why people say not to trust random WiFi networks.

  • source
  • parent
  • [–] 2 points 18 hours ago (1 child)

    For option 1, since you are the device, you just need to use the TLS certificate from the session, which you can capture otherwise. This is just the first source I could find: https://support.atlassian.com/atlassian-knowledge-base/kb/how-to-capture-http-traffic-using-wireshark-fiddler-or-tcpdump/

    Decrypt with SSLKEYLOGFILE

    Wireshark has the ability to use SSLKEYLOGFILE to decrypt https traffic. This file is a feature provided by the web browser. When a Web Browser is configured to create and use this file all of the encryption keys created for that session are logged. This allows Wireshark to decrypt the traffic. If you supply SSLKEYLOGFILE and a pcap file that were taken at the same time, wireshark will show you all of the web traffic.

  • source
  • parent
  • hideshow 1 child comment
  • load more comments (-1 replies)
  • [–] 20 points 23 hours ago (3 children)

    Illuminating, thanks for the info.

    But do consider. From a security standpoint, do you want your security software company running a bunch of local ML on your phone? Even if it were being run by a stand-up, security-first guy who didn't tell them to fuck all delays/testing and charge ahead, do you want them to put a bunch of data processing on your device? Do you want them all over your text and images, even locally, making an index that can be exfiltrated?

    Even if you completely remove the company's previous actions, I don't want that included in my VPN / Password / Document Storage app. I don't really even want that on my phone, even if it's provided by the Operating System.

    They use ML to read and summarize data, it's inches away from selling telemetry info about you.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 2 points 11 hours ago

    So what should they do in your opinion? They remove the OCR scanning, and then people will say "I'm never moving from Google Docs to Proton, they don't even have basic features like OCR scanning".

    Genuine question, I do agree with the concerns, and I also see how this is a basic feature any sensible office suite should have.

  • source
  • parent
  • [–] 10 points 23 hours ago (5 children)

    Any code from google is either a problem or a future problem. And before you say "what about x?" my answer will be "nope".

  • source
  • parent
  • hideshow 5 child comments
  • [–] 2 points 11 hours ago* (2 children)

    Ok, then you don't use any Android apps? Because the majority of code running in any Android all is Google written.

    I don't disagree it's a problem, but if the stance is that any Google code is bad, then whether they have this or not, one shouldn't use the app, and the Android phone in the first place, anyway.

  • source
  • parent
  • hideshow 2 child comments
  • [–] -2 points 23 hours ago (1 child)

    This isn't at all an implementation of OCR for the purposes of using a phone as a scanner. This is putting the foundations in to allow AI to utilize the contents of your e-mails.

    That's a huge privacy nightmare...

    Ideally the only people reading e-mails should be the senders and the recipients. Instead you also have this thing--doesn't matter if it's offline or online. It's a piece of software trying to contextualize private communications.

  • source
  • parent
  • hideshow 1 child comment