Schneier on Security

Clive Robinson • May 7, 2013 5:17 PM

Perhaps the question that should be asked is,

What do they need to store?

Followed by the question,

Why?

That is how much fidelity is required for each recording. That of basic inteligability or that for presentation in a court of law?

For the over one hundred languages spoken on a fairly regular basis the idea of “speach to text” is a bit of a misnomer, it’s more likely to be “speach to phonem” along with other bits of information indicating the way the phonems are said such as attack and decay times etc.

We know the likes of the NSA where the inventors of many of the basic codecs used to reduce high bandwidth speach several decades ago so perhaps it would not be two surprising if they had systems capable of analysing the call in progress and using some form of variable rate compression based on the inteligability of the phonems, and likewise deduce the language(s) being spoken etc.

We know that plain spoken text can be compressed to as little as 300bits/sec if your interest is spoken content and basic emotion, not speaker identification or background information/noise. In some cases a good deal less.

We also know that the NSA have in the past been the world leaders in storage of data.

But even with good compression algorithms there is still very definately a quality-v-storage issue. But there is potentialy a way around this and it’s to make the compression / quality time based. That is the older the recording the less use it is for legal or other purposes so the lower the quality of recording required and the more it can be compressed to what is essentialy just a transcript. The simplest way to do this would be in effect to make all the various quality levels of recordings at the same time and then at certain time intervals just delet the higher quality recordings unless flags etc had been set.

We also have reason to believe that all the above is currently achivable to a very high number of calls.

However as others above have noted there are issues around getting at the actuall calls.

Lets assume for argument sake (as in the UK) the bulk of calls are routed not along the old System X / Signaling System Seven but TCP/IP. This means that a call can be routed around any capacity issues and a local call of geographicaly just a mile or so could unnoticably to the call participents be routed around upto 100 miles within the network.

This gives an approximate range limit of 50miles to a “Tap Point” which for the bulk of US Citizens is well within the range to their nearest major city.

Whilst this would on the surface of it be a major problem, and would have been a few years ago it’s well within the capability of systems you can go and by from the likes of Cisco.

But whilst this may be possible at the lower levels what of the exchanges etc that sit above these layers. How do you get to corelate the actual switched conversation at the CO level with the individual packets on the underlying TCP/IP level. To do this would require specialised equipment whilst not impossible to make etc it would require a considerable number of specialists.

The reason for this is that although protocols are supposadly standard and open, in practice they may not. That is whilst an exchange manufacturer will support open standrds on it’s ports, the manufacture is likely to have “augmented” the standards when it’s equipment is talking to it’s equipment not that from another manufacturer (this has been seen in the past). Further as has been pointed out in recent times a lot of the exchange equipment used in the US and most other places around the globe are actually manufactured by just a small handfull of manufactures atleat two of which are Chinese.

So whilst it’s technicaly possible to do the mass surveillance of phone conversations I suspect it’s only upto a certain percentage of citizens around the likes of major cities etc.

The real question is if it is going on how much manpower is required and if in the hundred or above how come nobody has realy had a fit of concience and broken the secrecy.

Clive Robinson • May 8, 2013 4:58 PM

@ Bruce,

There are two possible ways to do this. The first is to collect, compress transport, and store. The second is to collect, convert to text, transport and store.

As you noted above there is more than one way “to skin a cat” on this, and I actually think the number of ways is considerably more than two, and importantly will evolve with time and technology as you cannot put a system such as this in place over night nore upgrade it all over night.

It appears the basic math on storage checks out and appears reasonable. However there is a lot more to it than just the basics of storing the data at an assumed bit rate. As I observed above the data compression / quality can be aged from high quality down to text transcript over time. And further can even be deleted with time and other considerations (you most likely don’t need to keep a recording of a seven year old describing their birthday party to granny for more than a few years and only as a text transcript after a few months). Thus with time primary storage is recovered, in essence the storage would be multilevel and most would end up off line fairly quickly. This is much like the multi layer memory behaviour on computers with CPU register, CPU cache, Main RAM, Hard Drive, backup tape, with access speed varying by 10^16 (ie pico sec down to 30day backup).

Thus we need to consider not just hard drive storage but high speed RAM and tape tractor systems. When you do this other factors arise which have important considerations.

For instance, as has been noted there is a power issue, having lots of hard disks with their platters spining almost continuously is going to make the lights dim in most major cities and towns if you tried putting them all in one place (which is perhaps unlikely on the “all the eggs in one basket” principle). When you start considering various storage technologies you end up with quite a complex power calculation per bit of storage. My guess is that hard drives will not actually be much more than a (very) high capacity buffer between RAM used for processing / searching and more permanent storage such as tape tractor farms.

But overall power and power per bit are just two of many considerations when you are looking at an operation that big, especialy when many things such as the environment are beyond your control.

Almost your first assumption would be “no single thing is 100% available / reliable / upgradeable” thus you would look at distributing it as a first measure (on the “divide and conquer” as well as “eggs in basket” principles).

Mind you even with carefull planning it can still go wrong. For instanceGoogle that does this to a certain extent has had and will continue to have occassional outages, and so for that matter will the telephone network. The important thing is “routing around a problem” which is considerably easier in a distributed de-centralised system.

However being widely distributed appears to fly in the face of fast data processing for searching etc. Which I guess is what Utha is all about.

But… it’s important to remember that searching and collection are very different in their charecteristics. Collection is a real time activity which needs nigh on 100% reliability and down times to be less than the network being monitored. Searching is far from real time and down times of an hour or two are acceptable if they occure infrequently.

Thus if it were down to me I certainly would be thinking seperate systems for collection and searching with a “store and forward” design where the collecting nodes stored in real time and forwarded as and when they could. It would also be helpfull if the collection heads could do some pre-prossessing over and above just compression (more of which later).

This also has the advantage that new technology needs only to go in at a modest rate with a reasonable life time (say five years).

But to be honest I think Utha is almost a “stop gap” measure, because the reality is a de-centralised system is going to be the best option in the long term partly because as we’ve known for quite some time centralised systems are “targets of opportunity” and partly because technology in the commercial sector is outgrowing user abilities.

And it is this last factor people need to think about carefully because the ultimatly decentralised system would run almost entirely on the end user device or “smart phone”.

People complain about the current voice recognition on their smart phones but it is noticably improving and will do so more quickly with time (because users despite their grumbles see it as useful technology). One advantage of having it on the end user device is it can learn the end user and thus get more recognition for less CPU power than a system that cannot learn individual users.

Thus it won’t be to long before your phone can speach to text everything you say. And as was seen with a commercial piece of software (CarrierIQ) the networks will quite happily do an end run around user security and send all SMS and other activities in the clear across the internet and thus do a big chunk of the NSA’s job for them…

Clive Robinson • May 10, 2013 4:25 AM

@ Buck,

For argument’s sake let’s say we do have 100% accuracy for speech to text. How does one ascribe metadata for code speak? Inside jokes? Sarcasm? Good luck!

You don’t need to.

Firstly have a look at how Google does language convertion, it does not use a dictionary as you or I would do. It uses the statistics of the data it has with regards the vast tract of various languages. In effect it translates sentances within the context of the document and other documents it has. The result is it does understand slang, jokes and various other nuances of day to day usage of a language.

The problem with people tryying to convay other information is that it forms styalised speach at variance to the norm. This changes it’s statistics just like any other form of steneography and there are known ways of detecting it as such.

As for speach to text not being 100% reliable it realy does not need to be better than around 50% to get usefull results when you are analysing not as words but sentances and paragraphs within a general context.

The reason being you need to turn the problem on it’s head, they are not looking to find a needle in the hay, but how to identify hay so it can be thrown away. And this can be done by successive filters. Thus the filters can be very crude rejecting only one or two percent at each stage but still producing high quality results (to see why go and look up how a uranium enrichment process works where each centrefuge only provides a fractional percentage of enrichment).