Silk Road forums

Quote from: masterblaster on October 27, 2012, 03:44 pm

Quote

For better voice quality,
Alice can use low-latency anonymity networks such as Tor [3] and JAP [4].

Something tells me these researchers dont know what they're talking about.

VOIP over TCP(the only protocol tor supports) is problematic because TCP will keep retrying old packets and delaying new ones. In a voice conversation if a bit of info was lost then it is better to just play what is coming next and forgetting about it, that is why UDP is used.

Real time protocols don't really play well with anonymizing networks. Far better to just use chat.

At this moment I am working on a chat client/server that opens a tor hidden service to facilitate secure communication.

In this thesis, we propose a class of passive traffic analysis attacks to com-
promise privacy of Skype calls and SIP-Based VoIP calls. The proposed attacks
are based on application-level features extracted from VoIP call traces. The proposed
attacks are evaluated by extensive experiments over different types of networks includ-
ing commercialized anonymity networks and our campus network. The experiments
show that the proposed traffic analysis attacks can detect speeches and speakers of
SIP based VoIP calls with 0.65 and 0.32 detection rate respectively, about 70-fold and
35-fold improvement over random guess. For Skype calls, the speech detection rate
and speaker detection rate are 0.33 and 0.44, about 30-fold and 15-fold improvement
over random guess. Countermeasures are proposed to mitigate the proposed traffic
analysis attacks by camouflaging. The proposed countermeasures can largely mitigate
the traffic analysis attacks and does not cause significant degradation on quality of
VoIP calls.

The proposed traffic analysis attacks exploit silence suppression. Different
people have different talk patterns in terms of talk spurts and silence gaps. For
example, some persons speak very fast with only a couple of short silence gaps while
some speak with long silence gaps. As shown in Figure 1, an eavesdropper can learn
a speaker’s talk pattern from packet timing. Based on talk patterns learned from
packet timing, the proposed traffic analysis attacks can detect speeches or speakers
of encrypted VoIP calls with high accuracy.

The typical attack scenario focused in this chapter is as follows: An adversary
may want to detect whether the target speaker, say Alice, is communicating with
Bob now or not based on the previous encrypted VoIP calls made by Alice. The
previous calls may use different codecs than the one Alice using now. The adversary
may collect VoIP packets at any point on the path from Alice to Bob and may also
want to detect the content of the conversation, such as a partial speech in previous
calls.
Comparing with previous researches, the proposed attacks do not require si-
multaneous access to both sides of the links connected to Alice and Bob. Traces
of calls used in detection can be collected at different time and in different network
environment and these calls possibly made with different codecs.

In this chapter, we also assume Alice uses mass-market VoIP services to com-
municate with Bob as shown in Figure 2. In other words, we assume SIP and RTP
are used as the signaling protocol and the transport protocol respectively. To protect
confidentiality of her VoIP calls, we assume Alice encrypts her VoIP packets by using
secure versions of the RTP protocol such as SRTP [2] and ZRTP used in Zphone [1].
To better protect privacy of her calls, we assume Alice routes these encrypted
VoIP calls through anonymity networks as shown in Figure 2. For better voice quality,
Alice can use low-latency anonymity networks such as Tor [3] and JAP [4].

We focus on passive attacks in this thesis. In other words, the attacks launched
by the adversary will not disturb existing network traffic. In comparison with active
attacks, the proposed attacks are harder to detect. We assume that the adversary
11
only has access to the links directly connected to participants of VoIP calls. This
assumption is widely used in traffic analysis attacks such as attacks on anonymity
networks and tracing VoIP calls [5, 8, 29, 30]. We do not assume the adversary as a
global attacker because rerouting techniques used in anonymity networks make global
attacks too costly to be practical. The threat model is weaker than threaten models
defined for traditional privacy-related traffic analysis attacks: The threat model does
not require simultaneous access to the links connected to participants of a VoIP
call which may not be feasible for international VoIP calls. Instead we assume the
adversary can collect traces of VoIP calls made by Alice in advance and use these
collected traces to detect whether Alice is a participant in the VoIP conversation
of interest. Our model is similar as the model for identifying a human being by
fingerprints: Fingerprints of human beings are collected in advance through driver
license applications. To identify a specific person, the fingerprint of interest such
as a fingerprint in a crime scene will be compared against the person’s fingerprints
collected in advance.
The threat model assumes the detections are based on different VoIP calls. So
the speaker identification should also be independent of the voice content of VoIP
calls.

The proposed traffic analysis attacks are based on packet timing information.
As described in Section 3.1.2, silence suppression enables adversaries to recover talk
patterns in terms of talk spurts and silence gaps from packet timing. Adversaries can
create a Hidden Markov Model (HMM) to model Alice’s talk pattern recovered from
SIP-Based encrypted VoIP calls made by her. When adversaries want to determine
which SIP-Based encrypted VoIP call is made by Alice, adversaries can check talk
patterns recovered from the call of interest against Alice’s model.
The proposed attacks can be divided into two phases: the training phase and
the detection phase as shown in Figure 3. The two steps in the training phase are
feature extraction and HMMs training. The detection phase consists of three steps:
feature extraction, speech detection or speaker detection, and intersection attack.
The last step, intersection attack, is optional. We describe the details of each step
below.


The input and output of the feature extraction step are raw traces of VoIP calls
and feature vectors respectively. The feature vector used in the proposed attacks is
where n is the length of a feature vector, tsi and sgj denote the length of the ith talk
spurt and the jth silence gap respectively.
Talk spurts and silence gaps are differentiated by a silence threshold Tsilence : If
an inter-packet time is larger than the threshold, the inter-packet time is declared as
a silence gap. Otherwise the inter-packet time is declared as a part of one talk spurt.
Obviously the threshold Tsilence is critical to overall detection performance. We
did initial experiments to investigate the suitable range of the threshold for detection:
We feed voice signals to VoIP clients and collect VoIP packets generated by VoIP
clients. Different values of the threshold Tsilence are used to determine silence gaps.
Actual silence gaps can be found by checking marker bits in RTP packets which
indicate the start of talk spurts2 . We evaluate a value of the threshold by two metrics:
false positive rate and false negative rate. False positive rate is the fraction of talk
spurts that were erroneously declared as silence gaps. False negative rate is the
fraction of silences gaps that are erroneously declared as talk spurts. The experiment
results with different codecs3 are shown in Figure 4.
We can observe that for a wide range of the threshold Tsilence , both the false
positive rate and the false negative rate are low: When Tsilence is larger than 70ms,
the false positive rate is below 10% for all the codecs. The false negative rate is below
20% when Tsilence is less than 100ms. The wide range is because of the big difference
between inter-packet time of silence gaps and inter-packet time of talk spurts: Silence
gaps are in order of seconds. Inter-packet time during talk spurts is usually around
packetization delay of 20ms or 30ms for most codecs.
We can also observe that increasing the threshold Tsilence decreases the false
positive rate and increases the false negative rate. The changes in these two rates
are again because the inter-packet time during silence gaps is larger than inter-packet
time during talk spurts.

A big challenge in feature extraction is to filter out noise caused by random net-
work delays in silence tests. Random network delays can cause errors in silence tests
especially at receiving side: Because of random packet delays, inter-packet time dur-
ing talk spurts can become larger than the threshold Tsilence . The main idea of filtering
noise in silence tests is to determine a silence gap based on N successive inter-packet
intervals instead of one inter-packet interval. The silence test with filtering techniques
works as follows: If one inter-packet interval is larger than the threshold Tsilence , we
declare a new silence gap only when none of the following ⌊
Tsilence
packetization delay
⌋ − 14
inter-packet intervals are shorter than a threshold Tspurt, used to filter out long inter-
packet intervals caused by network delays. The rationale behind the new silence tests
method is that: If an inter-packet interval is erroneously declared as a silence gap
because network delays increase the length of the inter-packet interval, then following
inter-packet intervals must likely be shorter than normal inter-packet intervals during
talk spurts. The new silence tests can improve silence detection performance in terms
of the false positive rate. The filtering does not focus on false negative errors because:
(a) The false negative rate changes very little when Tsilence changes. (b) We take into
account false negative errors in choices of HMM structures.

Our experiments clearly show that the proposed traffic analysis attacks can
greatly compromise privacy of VoIP calls. The detection rates for speech detection
and speaker detection are 70-fold and 35-fold improvement over random guess. Higher
detection rate can be achieved with more training traces.
Comparable detection performances are achieved for both traces collected by
sending side and receiving side. It is an indication that when the threshold is large
enough, feature extracted in the proposed attacks are largely independent of network
dynamics.
The framework proposed in this chapter, including extracting application-level
features from network traffic traces and statistical analysis of extracted application-
level feature by HMMs, can be potentially used to infer other sensitive information
at application level. For example, the framework can be potentially used to detect
speaker’s emotion during a call suppose the speaker’s talk behavior can change signif-
icantly when the speaker’s mood changes. The framework may also be used to detect
different types of speeches such as seminar talk, conversation between two parties,
and classroom discussion. One of our future works is to explore the potential of the
framework experimentally and theoretically.

Liberty,

I see your point about *government* just becoming the default 'response' for everyone as in  "I dunno, the government should just take care of it ... ".   I get it; nothing should have that much power over everyone, lest we all become so reliant on it that we can't imagine life without it. I get it bro. That kind of logic is sound.

BUT - in the meantime, who's gonna build a practical transportation system of roads and infrastructure? Who will setup reliable schools nationwide? For God's sake man - who will subsidize my electrical bill ??!!
A series of loose knit entrepreneurs who believe in nothing other than making a profit - OR - some kind or elected representative from my neighborhood ( HINT - you get a say in what happens with one of those choices.)

Now don't get me wrong - I'm not for everything being government run, in my mind the ONLY things government should be involved in are roads, education, hospitals, a judiciary and a BASIC welfare safely net so that those of us who don't succeed will not starve.
In other words - only things that work for the clear benefit of us all, everything else should be left to private industry to sort out.

In my world, we're not afraid of our government because we control it; we keep it in check constantly by participating in the communities where we live, making sure that our opinions are expressed by the freedoms we are guaranteed, and we really are guaranteed them! Publish any opinion you want in any media outlet and you will never get arrested for it (as long as your not preaching hate, death and destruction Duhh !!)

Pretty interesting and actually scary vision of revolution. First we get DEA and other goverment 'opressors' then who? Anyone who disagrees with us doing drugs? Your parents, priests?

And as far as I know in most western countries you can't punish someone after the law changes for something that he did according to the earlier law(unless it is genocide or homicide).

I'm from post communistic country, where more than half of population belonged to a party, worked as secret police or worked with secret police. Most of them did so just to get along without problems or fear, some of them because they actually believed in the system. And here comes 90s when communism goes down the shitter and we have to do something with all these people who helped to oppress others. There were couple of ways to do it(Germany and Czech Republic did it pretty well) my country chose the worst way(mostly due to heads of opposition being secret service themselves) so now the people who were in power are still in power and in the end not even people who gave orders to shoot at protestors were held responsible. Yet the question was/is who, how and why should we punish for oppression when more than half of the people in the country were responsible for it.

Anyways funny thing is that someone here claims that SR is being anarchist, with escrow, moderation, ratings etc. it is far from being an anarchy. Most organized crime(as in going against the law not morally/philosophically crime) is far from being an anarchy.