While I would love deeper technical detail on Face ID, this is an excellent overview. It sounds like one of the best-engineered biometric systems for mass deployment I have ever seen. Randomized per-device dot patterns? That's amazing. I wonder how dissimilar the projected dot patterns are from one another. For example, if I got dot patterns from 100 iPhone X units, what are the chances I would see the same pattern twice? What about the chances of two having 95% of the dots in the same place?
I wonder what attacks might be possible on the depth mapping. For example, is the depth map built in the TrueDepth module itself? It sounds like it is, and then it is signed and shipped off to the Secure Enclave. If so, forging depth map data could be possible, but would be extremely difficult.
I don't believe it's Apple's intent at least for now to claim that face ID is the end-all and be-all and no one can get into your phone if your face doesn't match. Even within the white paper Apple advises that if you have siblings who closely resemble you or if you have children under the age of 13 that a passcode should be instead because the chances of misidentification is higher. My guess is that's why a passcode needs to be entered at least every 4 hours if your phone goes unused.
With Apple making a special point of setting up a system for sorting face ID issues should make it obvious that they are still fine-tuning this. Give it another year and time for Apple to tweak it
The mask-rejection network is probably the best place to attack for the foreseeable future. It would be very interesting to see the false-positive and false-negative rates of that recognizer. ... What if there's a face, but that person isn't looking at the phone (say, a coworker bumps the phone while looking at something else)?
A neural network in the Secure Enclave is specially trained to identify masks and photos. The reflection of the infrared from the skin and from the mask should certainly differ and the neural network is certainly smart to detect that. There may be other differences to be never disclosed, such as some patterns or values specific to all masks the mask producers would never know. The mask must be built very cleverly to spoof the neural network. But for that first the neural network must be reverse engineered, which is impossible since it is enclosed in the secure enclave.
The detection of a face is not enough, a directed gaze too is needed. If there exists both then this is a false face and it remains locked. Five false faces like that and it switches to password mode.
"Users will also be required to use the passcode if it hasn't been used to unlock the iPhone X in the last 156 hours and if Face ID has not been used successfully in the last four hours."
The above is from the article. If I read that right, and please tell me if you think I'm wrong, but isn't that saying if you don't use your phone for four hours you need the passcode? That's ridiculous. Right now I think it's 48 hours then you have to use the passcode if you haven't used the Touch ID.
I read an AND in that sentence, which implies both conditions must be true. So no worries; it’s not as onerous as it would be if the sentence contained an OR.
For Apple Pay purchases in stores, users have to confirm intent to pay by a double-tap of the side button, followed by a Face ID authentication, before placing the iPhone X near the contactless reader. Users will have to reauthenticate with Face ID if they change a different Apple Pay payment method, but will not need to tap the button again.
"Confirm intent to pay"?
Currently, when paying with Apple Pay, I simply hold my iPhone near the payment terminal and my credit card is automatically displayed, I confirm my touching my finger to the Home button and that's it.
With FaceID will I have to hold the iPhone near the payment terminal, then double-tap the side button, then authenticate with my face and then hold the iPhone to the reader again? This seems like a slightly more cumbersome procedure. I suppose I could double-tap the side button while pulling the iPhone from my pocket, authenticate and then hold to the reader, but will that work?
Overall, it doesn't matter terribly as I usually use my Apple Watch for Apple Pay. However, twice lately (at Big Y) my Apple Watch hasn't been successful and I've had to use my iPhone as a backup.
Sounds like sh**. I’ve been noticing TouchID more now and yes it’s easier to just look at the screen for locked apps but Apple Pay seems like a pain in comparison.
I use ApplePay all the time with my iPhone6. I have multiple cards enrolled but usually just use the default one. To save time at the checkout, I always double click the home button and pre-authenticate with my finger print whilst the staff member is ringing up my stuff. I just need to hold near the terminal and it's done. I can't see how this is any different to the FaceID/ apple watch method.
For Apple Pay purchases in stores, users have to confirm intent to pay by a double-tap of the side button, followed by a Face ID authentication, before placing the iPhone X near the contactless reader. Users will have to reauthenticate with Face ID if they change a different Apple Pay payment method, but will not need to tap the button again.
"Confirm intent to pay"?
Currently, when paying with Apple Pay, I simply hold my iPhone near the payment terminal and my credit card is automatically displayed, I confirm my touching my finger to the Home button and that's it.
With FaceID will I have to hold the iPhone near the payment terminal, then double-tap the side button, then authenticate with my face and then hold the iPhone to the reader again? This seems like a slightly more cumbersome procedure. I suppose I could double-tap the side button while pulling the iPhone from my pocket, authenticate and then hold to the reader, but will that work?
Overall, it doesn't matter terribly as I usually use my Apple Watch for Apple Pay. However, twice lately (at Big Y) my Apple Watch hasn't been successful and I've had to use my iPhone as a backup.
I suppose the "double tap" of the side button is a security measurement that you are indeed wanted to make an ApplePay payment, that you do not just accidentally press your phone to the terminal. If you think about it, the breakdown would only be: - The merchant's payment reader is ready - You double tap the side button while moving the phone over the terminal (since the phone is facing you, it's already done its part to authenticate you by FaceID, no additional action is required) - Done.
Comparing to the current method, I don't see it's more cumbersome than to move your finger over the TouchID while moving the phone over the terminal, similar to step 2 above?
While I would love deeper technical detail on Face ID, this is an excellent overview. It sounds like one of the best-engineered biometric systems for mass deployment I have ever seen. Randomized per-device dot patterns? That's amazing. I wonder how dissimilar the projected dot patterns are from one another. For example, if I got dot patterns from 100 iPhone X units, what are the chances I would see the same pattern twice? What about the chances of two having 95% of the dots in the same place?
I wonder what attacks might be possible on the depth mapping. For example, is the depth map built in the TrueDepth module itself? It sounds like it is, and then it is signed and shipped off to the Secure Enclave. If so, forging depth map data could be possible, but would be extremely difficult.
I don't believe it's Apple's intent at least for now to claim that face ID is the end-all and be-all and no one can get into your phone if your face doesn't match. Even within the white paper Apple advises that if you have siblings who closely resemble you or if you have children under the age of 13 that a passcode should be instead because the chances of misidentification is higher. My guess is that's why a passcode needs to be entered at least every 4 hours if your phone goes unused.
With Apple making a special point of setting up a system for sorting face ID issues should make it obvious that they are still fine-tuning this. Give it another year and time for Apple to tweak it
I think you are misreading the requirements. The four-hour clock only starts *after* the 156-hour clock runs out.
I'm thinking more about attacks specifically to pass off false data as true. The dot projector pattern uniqueness is only a partial defense due to the birthday paradox. Still, more layers are always good.
The mask-rejection network is probably the best place to attack for the foreseeable future. It would be very interesting to see the false-positive and false-negative rates of that recognizer.
Attacks from within the system are probably limited. To the best of my knowledge, nobody has arbitrary code execution on the Touch ID processor. Face ID will probably be similar in that regard. They just don't need to receive data from anything in the rest of the phone, so input opportunities are limited. Getting the TrueDepth module to execute arbitrary code would almost certainly be required to get it to sign forged depth data.
I am also very curious about what exactly burns a face recognition attempt. For example, if the phone wakes up in my pocket and sees nothing, does that count as a failure to recognize a face? If it's sitting face-up on my desk and I tap the screen to wake it, but I'm way off to the side and the camera only sees the ceiling, does that count? What if there's a face, but that person isn't looking at the phone (say, a coworker bumps the phone while looking at something else)?
I’m gonna guess there’s a face detection process that loops first, with some timeout. Face detection is a light duty process compared to face recognition, and would return coordinates of any faces located within view of the camera to kick off the rest of the sensors involved in the face recognition process. So, yeah, a co-worker bumping the phone, or table it’s on, might kick off the face detection process, and that in turn might kick off the face recognition process, but then he or she, or someone else within view, would have to be looking at the phone before a negative attempt would be registered. This suggests to me something Apple could incorporate. A gender recognition process, to reject (not count as an attempted unlock) a person recognized as a different gender than the registered face but otherwise not identified as the registered user. This would allow your significant other, in the case where that person is a different gender, to be in tne scene, as might be the case when handing you your phone - a common scenario in homes, at east in mine - without consuming one of the limited number of recognition attempts.
"Users will also be required to use the passcode if it hasn't been used to unlock the iPhone X in the last 156 hours and if Face ID has not been used successfully in the last four hours."
The above is from the article. If I read that right, and please tell me if you think I'm wrong, but isn't that saying if you don't use your phone for four hours you need the passcode? That's ridiculous. Right now I think it's 48 hours then you have to use the passcode if you haven't used the Touch ID.
The second condition only applies if the first condition is met.
This means that you'll have to use your passcode to unlock the iPhone once a week even if none of the other conditions are met. The second condition - the 4 hours - means that the iPhone won't force you to enter the passcode (if it hasn't been entered in 6-1/2 days) while you're using it, e.g. in the middle of the day. It will wait until after you haven't used the iPhone for a while, e.g. while you were sleeping, to force you to make your once-a-week passcode entry.
Hmm, makes me wonder if this is the time when the iPhone will take advantage of the knowledge that it must be the phone’s owner unlocking the phone, because the password is being used. What better time to do a scan to add to the stored data any changes to the user’s face?
"Users will also be required to use the passcode if it hasn't been used to unlock the iPhone X in the last 156 hours and if Face ID has not been used successfully in the last four hours."
The above is from the article. If I read that right, and please tell me if you think I'm wrong, but isn't that saying if you don't use your phone for four hours you need the passcode? That's ridiculous. Right now I think it's 48 hours then you have to use the passcode if you haven't used the Touch ID.
I understand this to mean, I’d face id gets a false face then and isn’t used with
“Users will also be required to use the passcode if it hasn't been used to unlock the iPhone X in the last 156 hours and if Face ID has not been used successfully in the last four hours”
I see two separate clauses here.
I see 156 hours no pass code. Then I see if face id wasn’t used to SUCCESSFULLY unlock the iPhone in the last 4 hours. Then you’ll need your passcode.
The way it’s written could also mean that if you don’t use face id for 152 hours then decide to use face id to unlock your phone but fail then wait 4 hours to unlock your phone it will require a passcode. To me these like two separate things that the author just put in the same sentence.
In Settings > Wallet & Apple Pay, there is an option to bring up the wallet with a double-click on the home button. I don't know if it is on by default, but it is off on my phone and I do not remember changing it. With that, the flow would be like the Face ID flow:
1. Double-click the home button to bring up the wallet. If you leave your enrolled finger on it, it can scan your finger to verify identity. 2. Pick the card you want to use if it is different from the default. 3. Hold your phone to the terminal.
This last one you presented is the way I've always used TouchID. I don't even ever remember setting the "double click to open wallet" on my phone, though if it's not the default, I guess I must have at some point.
I'm not as dexterous with my left hand as some, so double clicking a side button rather than the Home button would be more cumbersome, but it's a moot point. I won't be getting a new phone this year anyway.
The 156 hour/4 hour statement is confusing - not sure what to make of that. Either way, I probably end up entering my pass code on my 6s at least once a day anyway, so it's really. Not that big of a deal. The accuracy and reliability of Face ID will be the bigger issue.
I honestly can't see what the big deal is with the 'privacy' concerns of Face ID. At a basic level, the technology is very similar to Touch ID, and everyone is fine with that. Furthermore, I consider a fingerprint a much more 'private' identifier than my face, so why is everyone suddenly all worked up? If you are truely concerned with Face ID or Touch ID, just use a passcode. None of these systems are required, people!
As an aside, I work at a hospital, and the Pyxis drug dispensing machines that the hospital uses require you to use a fingerprint - but no one has complained there, not even the union, and we have absolutely no information about how our fingerprints are scanned, stored, transmitted, sent to the NSA....
While I would love deeper technical detail on Face ID, this is an excellent overview. It sounds like one of the best-engineered biometric systems for mass deployment I have ever seen. Randomized per-device dot patterns? That's amazing. I wonder how dissimilar the projected dot patterns are from one another. For example, if I got dot patterns from 100 iPhone X units, what are the chances I would see the same pattern twice? What about the chances of two having 95% of the dots in the same place?
I wonder what attacks might be possible on the depth mapping. For example, is the depth map built in the TrueDepth module itself? It sounds like it is, and then it is signed and shipped off to the Secure Enclave. If so, forging depth map data could be possible, but would be extremely difficult.
I don't believe it's Apple's intent at least for now to claim that face ID is the end-all and be-all and no one can get into your phone if your face doesn't match. Even within the white paper Apple advises that if you have siblings who closely resemble you or if you have children under the age of 13 that a passcode should be instead because the chances of misidentification is higher. My guess is that's why a passcode needs to be entered at least every 4 hours if your phone goes unused.
With Apple making a special point of setting up a system for sorting face ID issues should make it obvious that they are still fine-tuning this. Give it another year and time for Apple to tweak it
The mask-rejection network is probably the best place to attack for the foreseeable future. It would be very interesting to see the false-positive and false-negative rates of that recognizer. ... What if there's a face, but that person isn't looking at the phone (say, a coworker bumps the phone while looking at something else)?
A neural network in the Secure Enclave is specially trained to identify masks and photos. The reflection of the infrared from the skin and from the mask should certainly differ and the neural network is certainly smart to detect that. There may be other differences to be never disclosed, such as some patterns or values specific to all masks the mask producers would never know. The mask must be built very cleverly to spoof the neural network. But for that first the neural network must be reverse engineered, which is impossible since it is enclosed in the secure enclave.
The detection of a face is not enough, a directed gaze too is needed. If there exists both then this is a false face and it remains locked. Five false faces like that and it switches to password mode.
Well yes, the existence of the mask-rejection network is why I talked about the mask-rejection network. ;-)
IR reflectivity almost certainly is not a differentiator. Apple definitely is not using a far-infrared camera here. You can cover a face in makeup and cover a mask in exactly the same makeup, and their reflectivity within the spectrum Apple is using will be identical. I suspect it is instead capturing the face over time and watching for movement. Skin over muscles moves differently from a prosthetic over skin over muscles.
Of course, due to the training process, there is no real way to tell (even for the developers) what exactly it is using to tell the difference. That is why I am more curious about the raw false-positive and false-negative rates.
While I would love deeper technical detail on Face ID, this is an excellent overview. It sounds like one of the best-engineered biometric systems for mass deployment I have ever seen. Randomized per-device dot patterns? That's amazing. I wonder how dissimilar the projected dot patterns are from one another. For example, if I got dot patterns from 100 iPhone X units, what are the chances I would see the same pattern twice? What about the chances of two having 95% of the dots in the same place?
I wonder what attacks might be possible on the depth mapping. For example, is the depth map built in the TrueDepth module itself? It sounds like it is, and then it is signed and shipped off to the Secure Enclave. If so, forging depth map data could be possible, but would be extremely difficult.
I don't believe it's Apple's intent at least for now to claim that face ID is the end-all and be-all and no one can get into your phone if your face doesn't match. Even within the white paper Apple advises that if you have siblings who closely resemble you or if you have children under the age of 13 that a passcode should be instead because the chances of misidentification is higher. My guess is that's why a passcode needs to be entered at least every 4 hours if your phone goes unused.
With Apple making a special point of setting up a system for sorting face ID issues should make it obvious that they are still fine-tuning this. Give it another year and time for Apple to tweak it
The mask-rejection network is probably the best place to attack for the foreseeable future. It would be very interesting to see the false-positive and false-negative rates of that recognizer. ... What if there's a face, but that person isn't looking at the phone (say, a coworker bumps the phone while looking at something else)?
A neural network in the Secure Enclave is specially trained to identify masks and photos. The reflection of the infrared from the skin and from the mask should certainly differ and the neural network is certainly smart to detect that. There may be other differences to be never disclosed, such as some patterns or values specific to all masks the mask producers would never know. The mask must be built very cleverly to spoof the neural network. But for that first the neural network must be reverse engineered, which is impossible since it is enclosed in the secure enclave.
The detection of a face is not enough, a directed gaze too is needed. If there exists both then this is a false face and it remains locked. Five false faces like that and it switches to password mode.
Well yes, the existence of the mask-rejection network is why I talked about the mask-rejection network. ;-)
IR reflectivity almost certainly is not a differentiator. Apple definitely is not using a far-infrared camera here. You can cover a face in makeup and cover a mask in exactly the same makeup, and their reflectivity within the spectrum Apple is using will be identical. I suspect it is instead capturing the face over time and watching for movement. Skin over muscles moves differently from a prosthetic over skin over muscles.
Of course, due to the training process, there is no real way to tell (even for the developers) what exactly it is using to tell the difference. That is why I am more curious about the raw false-positive and false-negative rates.
While I would love deeper technical detail on Face ID, this is an excellent overview. It sounds like one of the best-engineered biometric systems for mass deployment I have ever seen. Randomized per-device dot patterns? That's amazing. I wonder how dissimilar the projected dot patterns are from one another. For example, if I got dot patterns from 100 iPhone X units, what are the chances I would see the same pattern twice? What about the chances of two having 95% of the dots in the same place?
I wonder what attacks might be possible on the depth mapping. For example, is the depth map built in the TrueDepth module itself? It sounds like it is, and then it is signed and shipped off to the Secure Enclave. If so, forging depth map data could be possible, but would be extremely difficult.
I don't believe it's Apple's intent at least for now to claim that face ID is the end-all and be-all and no one can get into your phone if your face doesn't match. Even within the white paper Apple advises that if you have siblings who closely resemble you or if you have children under the age of 13 that a passcode should be instead because the chances of misidentification is higher. My guess is that's why a passcode needs to be entered at least every 4 hours if your phone goes unused.
With Apple making a special point of setting up a system for sorting face ID issues should make it obvious that they are still fine-tuning this. Give it another year and time for Apple to tweak it
The mask-rejection network is probably the best place to attack for the foreseeable future. It would be very interesting to see the false-positive and false-negative rates of that recognizer. ... What if there's a face, but that person isn't looking at the phone (say, a coworker bumps the phone while looking at something else)?
A neural network in the Secure Enclave is specially trained to identify masks and photos. The reflection of the infrared from the skin and from the mask should certainly differ and the neural network is certainly smart to detect that. There may be other differences to be never disclosed, such as some patterns or values specific to all masks the mask producers would never know. The mask must be built very cleverly to spoof the neural network. But for that first the neural network must be reverse engineered, which is impossible since it is enclosed in the secure enclave.
The detection of a face is not enough, a directed gaze too is needed. If there exists both then this is a false face and it remains locked. Five false faces like that and it switches to password mode.
Well yes, the existence of the mask-rejection network is why I talked about the mask-rejection network. ;-)
IR reflectivity almost certainly is not a differentiator. Apple definitely is not using a far-infrared camera here. You can cover a face in makeup and cover a mask in exactly the same makeup, and their reflectivity within the spectrum Apple is using will be identical. I suspect it is instead capturing the face over time and watching for movement. Skin over muscles moves differently from a prosthetic over skin over muscles.
Of course, due to the training process, there is no real way to tell (even for the developers) what exactly it is using to tell the difference. That is why I am more curious about the raw false-positive and false-negative rates.
Frequencies absorbed vs frequencies reflected...
Again, if you have a face with makeup and a mask with the same makeup on it, they will reflect and absorb infrared light in these wavelengths identically. This portion of the IR spectrum does not penetrate appreciably into skin.
Comments
The detection of a face is not enough, a directed gaze too is needed. If there exists both then this is a false face and it remains locked. Five false faces like that and it switches to password mode.
- The merchant's payment reader is ready
- You double tap the side button while moving the phone over the terminal
(since the phone is facing you, it's already done its part to authenticate you by FaceID, no additional action is required)
- Done.
Comparing to the current method, I don't see it's more cumbersome than to move your finger over the TouchID while moving the phone over the terminal, similar to step 2 above?
Hmm, makes me wonder if this is the time when the iPhone will take advantage of the knowledge that it must be the phone’s owner unlocking the phone, because the password is being used. What better time to do a scan to add to the stored data any changes to the user’s face?
I see two separate clauses here.
I see 156 hours no pass code. Then I see if face id wasn’t used to SUCCESSFULLY unlock the iPhone in the last 4 hours. Then you’ll need your passcode.
The way it’s written could also mean that if you don’t use face id for 152 hours then decide to use face id to unlock your phone but fail then wait 4 hours to unlock your phone it will require a passcode. To me these like two separate things that the author just put in the same sentence.
This last one you presented is the way I've always used TouchID. I don't even ever remember setting the "double click to open wallet" on my phone, though if it's not the default, I guess I must have at some point.
I'm not as dexterous with my left hand as some, so double clicking a side button rather than the Home button would be more cumbersome, but it's a moot point. I won't be getting a new phone this year anyway.
I honestly can't see what the big deal is with the 'privacy' concerns of Face ID. At a basic level, the technology is very similar to Touch ID, and everyone is fine with that. Furthermore, I consider a fingerprint a much more 'private' identifier than my face, so why is everyone suddenly all worked up? If you are truely concerned with Face ID or Touch ID, just use a passcode. None of these systems are required, people!
As an aside, I work at a hospital, and the Pyxis drug dispensing machines that the hospital uses require you to use a fingerprint - but no one has complained there, not even the union, and we have absolutely no information about how our fingerprints are scanned, stored, transmitted, sent to the NSA....
IR reflectivity almost certainly is not a differentiator. Apple definitely is not using a far-infrared camera here. You can cover a face in makeup and cover a mask in exactly the same makeup, and their reflectivity within the spectrum Apple is using will be identical. I suspect it is instead capturing the face over time and watching for movement. Skin over muscles moves differently from a prosthetic over skin over muscles.
Of course, due to the training process, there is no real way to tell (even for the developers) what exactly it is using to tell the difference. That is why I am more curious about the raw false-positive and false-negative rates.