Fast and Fair Handling of Multimedia CAPTCHA Flows

Multimedia CAPTCHAs play a crucial role in bot filtering policies for VoIP applications. Network flows carrying CAPTCHA content have a very small lifespan and require high transmission quality, making for special handling at transport layer. The present study introduces an analysis-derived, rapidly converging control mechanism for CAPTCHA flows. Instead of relying on generic heuristics, the novel scheme enables the flows to estimate their deviation from the equilibrium and adapt to it in a single step. Simulations demonstrate the high efficiency and convergence speed of the scheme, highlighting its unique fitness for the appointed task.


I. INTRODUCTION
Voice over IP services are envisioned as the future of telephony and online communications.However, thinning the borders between computer networks and telephony has its risks.Modern VoIP networks have to account for "bots", i.e. pieces of software that perform automated calls for purposes ranging from advertisement to denial of service [1].Call !ltering policies are employed in retaliation.An integral part of these policies are the CAPTCHAs, a form of reverse Turing test which is administered as a means of proving that the caller is indeed human.CAPTCHAs generally come in multimedia form and contain simple puzzles to be solved, involving text, video and sound recognition in the process [2].The proper management of CAPTCHA "ows at transport layer is the focus of the present study.
Performing CAPTCHA tests on users decreases their quality of experience.This situation can be aggravated further under presence of network issues.Audio/video problems on the received content could result into test failure, leading to caller blacklisting, blocking or banning, as per the enforced policy.Long delays in reception are not acceptable either, since the tests are administered with time restrictions.Using an error control-capable transport protocol (TCP variant) is in general mandatory.Thus, the problem can be reduced to proper bandwidth allocation for the CAPTCHA-conveying network "ows.
However, a unique attribute of a CAPTCHA "ow is its small size.The actual media content typically varies from a few KBytes to a few MBytes.Thus, a desirable attribute of a proper "ow management scheme is the fast convergence to equilibrium.The coexistence with "ows with larger lifespans (e.g.FTP, streaming) could pose an issue [3].A commonly used approach is then to employ separate router buffers for the short-lived "ows [4].The same outcome can also be accomplished by placing the "ows into virtual dedicated channels [5].Nonetheless, router buffer allocations, or dedicated bandwidth, are a scarce and valuable resource.The corresponding CAPTCHA "ows will eventually saturate the available buffers, causing congestion.The present work offers a novel way of CAPTCHA "ow management at transport layer.The study is motivated by the results of [6], which shows that the delay and jitter per "ow are minimized when all "ows are assigned their fair share of the bandwidth.A mechanism is presented that allows each "ow independently to estimate its deviation from its fairshare.More particularly, each "ow becomes aware of whether it has operated beyond, below or around its fairshare, which allows them to determine the next congestion control strategy for fast convergence to fairness.We then introduce the necessary window adjustment rules.According to the Fair-Share rule each "ow can estimate and rapidly operate at its fair-share.Simulation results con!rm that, in the presence of the Fair-Share rule, the system converges in one congestion cycle (one epoch) to equilibrium.

II. ANALYSIS
We assume the typical bottleneck topology of Fig. 1.A number ! !!!!!of "ows compete for the link and adjust their congestion windows, !!!!!, following the Additive Increase-Multiplicative Decrease rule [7].The system model is characterized by synchronous "ow noti!cation for congestion events, which is a classic assumption as well [8].We de!ne the instantaneous Throughput (!) of the !!! "ow at time t as: where R(t) is the time-variant round trip time, comprising the queuing delay !!!!! and a constant factor R o representing physical attributes: The flow windows undergo the following repeating procedure.Initially, all flows are in the additive increase stage.At the time the bandwidth is exhausted and the buffer has overflowed (cliff point) the network signals the senders to decrease their data rate.Consider a flow i that competes for the bw b bandwidth.After a successful delivery of a packet, it receives the relevant ACK from the receiver.Consider two ACKs that it receives between the system knee and cliff points [8], one at time t 1 and another at time t 2 .The throughput of the i th flow at time t 1 and t 2 are T i (t 1 ) and T i (t 2 ) respectively, and !! !!! !!is: denote the bandwidth and delay respectively.BS is the buffer size at the router.
where !! ! is defined as: Consequently from ( 3) and ( 4): Remark 1.If network resources were allocated equally among competing flows, they would all experience the same increase of their congestion window, and equal to !! !!! !!during the time period from t 1 to t 2 , defined as: Due to this remark, equation (5) becomes: might be either positive or negative.
Corollary 2. Equation ( 8) enables the i th flow to determine whether it operates below, beyond or at its fairshare as follows: • If !T(t 2 )<0, the i th flow at time t 1 has reached a rate above its fair-share, since , the rate of i th flow at time t 1 is below its fair-share, since • If !T(t 2 )=0, the i th flow has reached its fair-share, since Therefore, equation ( 8) establishes a criterion, based on which each flow independently can review its transmission rate and deduce whether it operates greedily, fairly or suffering mistreatment.Theorem 3. In order to operate at its fair-share, the flow should adjust its congestion window, according to equation: where Consider that the optimal congestion window values, that guarantee operation at its fair-share at times t A and t B are w C and w D respectively.Consequently, the straight line connecting points C and D is the fair-share line.According to (8) it will hold that T C =T D and, therefore: Since the additive increase factor is the same in both cases (A I =1), vector !" is parallel to vector !", and thus it holds that w C =w A +x and w D =w B +x. Consequently, from (10): Note that due to the different rate of incoming packets at the router during the optimal epoch, the flow f 1 might not have measurement samples for which R C =R A and R D =R B .However the fair-share line of the congestion window does not depend on these values, since the additive increase factor is static (e.g.unary).Hence, from (11) we derive: The optimal decrease of the congestion window at the end of j th epoch, should also guarantee that the system operates beyond the efficiency line.The latest is guaranteed as long as R knee' =R o .So, we seek to satisfy: From equations ( 12) and (13) we conclude: 2 nd case (operation below the fair-share, right part of Fig. 2).Since flow f 1 has consumed less resources than its fairshare, the throughput line of the flow is increasing and the optimal congestion window line, the fair-share line, is greater than the measured one.Similarly to before we conclude that: where 3 rd case (the flow has reached it fair-share).Via (8): Consequently, from (12) we deduce 0 u = and from (9): which is true, since the flow has operated at its fair-share line during the j th epoch.Conclusively, using Corollary 2 and Theorem 3 a flow can instantaneously set its congestion window to the fair-share value, without relying on simple, converging heuristics like the multiplicative decrease of the AIMD scheme.The employed active queue management, i.e. the "ow congestion noti!er mechanism, is the EGCN algorithm (Explicit Global Noti!er), which has been shown to promote fairness with regard to droptail-based approaches [6].The following simulations considered a varying number of "ows and present their time-variant congestion windows.The novel, Fair-share congestion management, denoted as TCP-FS, is compared to the standard TCP which relies on the Additive Increase-Multiplicative Decrease operation (AIMD) [7].
The following contention increase scenario is employed.Initially, two "ows with the same round-trip time, 50msec enter the system at 0sec and at 0:1sec respectively (Figure 3a).An additional "ow is introduced at 10sec with the same round-trip time (Figure 3b).Finally, at time 20sec, a fourth "ow enters the system with different round-trip time, 15msec, to study the impact of different RTTs (Figure 3c).
The behavior of the standard TCP is shown in Fig. 4. The "ows slowly converge to their fair shares after an interval of approximately 10sec.Notice that the standard lifespan of a CAPTCHA "ow can be equal or less, posing an issue of convergence potential overall.The behavior is expected, since the multiplicative decrease of the AIMD strategy is a heuristic that treats the network in a generic way.The same applies to newer variants, such as [10].Without explicit knowledge of the current deviation from equilibrium, the system is bound to converge in several iterations.For times 10-30 sec, TCP exhibits similar behavior as well.
Concerning the TCP-FS, it is shown that at the end of an epoch, where all co-existing "ows measure their Throughput samples, the "ows succeed in adjusting their window to their fair-share.This is con!rmed by the equally assigned congestion windows of Fig. 3a, 3b and 3c.Even when new "ows enter the system (e.g.flow 3 at time 10sec in Fig. 3b), all "ows adjust their rates to the new fair-share in just one epoch.Concerning the addition of the fourth "ow in Fig. 3c, it is observed that the Fairshare rule correctly assigns a smaller congestion window to it, in order to make up for the higher round-trip time.
Finally, Fig. 5 con!rms that the fair allocation of resources is achieved in conjunction with maintaining high levels of link utilization.The system operates between the knee and cliff points, and utilization is affected momentarily (no more than an epoch), which is the time required for all "ows to evaluate their new fairshare.

IV. CONCLUSION
A novel, rapidly converging congestion control scheme for the TCP protocol was presented.The new scheme promotes the fair handling of short-lived flows, an attribute of special interest in VoIP/CAPTCHA traffic.Analysis-derived metrics allow the competing flows to estimate their deviation from the equilibrium and converge in a single step.Simulations demonstrated the ability of the scheme to converge fast while maintaining high channel utilization.

Figure 1 .
Figure 1.A typical model of n users sharing a link.bw(*) and d(*) denote the bandwidth and delay respectively.BS is the buffer size at the router.

Figure 2 .
Figure 2. Flow operation above (left) and below (right) its fair-share.

Figure 3 .
Figure 3. Flow congestion windows for TCP-FS.All flows, new and existing, converge to their appropriate shares in just one epoch.

Figure 4 .
Figure 4. Congestion windows of flows handled by the classic TCP.The flows converge slowly to their fair shares at ! 10sec.However, a typical CAPTCHA or control flow can have equal or smaller lifespan.

Figure 5 .
Figure 5. Queue length of TCP-FS.The link remains highly utilized throughout the operation of the novel scheme.
We examine the cases defined by Corollary 2.SHORT PAPER FAST AND FAIR HANDLING OF MULTIMEDIA CAPTCHA FLOWS congestion window values (w), which guarantee flow f 1 operation at its fair-share during the j th epoch, is lower than the current value of operation.
A and ACK B , measures the round-trip time of packets, R A and R B , and calculates the corresponding values of throughput, T A, T B .Since flow f 1 has exceeded its fair-share, the throughput values during the period t knee -t cliff are decreasing and the iJIM -Volume 9, Issue 4, 2015