Category: Publications

Categories
Data Papers Publications

New paper “Auditing for Discrimination in Algorithms Delivering Job Ads” at TheWebConf 2021

We published a new paper “Auditing for Discrimination in Algorithms Delivering Job Ads” by Basileal Imana (University of Southern California), Aleksandra Korolova (University of Southern California) and John Heidemann (University of Southern California/ISI) at TheWebConf 2021 (WWW ’21).

From the abstract:

Skew in the delivery of real-world ads on Facebook (FB) but not LinkedIn (LI).
Comparison of ad delivery using “Reach” (R) and “Conversion” (C) campaign objectives on Facebook. There is skew for both cases but less skew for “Reach”.

Ad platforms such as Facebook, Google and LinkedIn promise value for advertisers through their targeted advertising. However, multiple studies have shown that ad delivery on such platforms can be skewed by gender or race due to hidden algorithmic optimization by the platforms, even when not requested by the advertisers. Building on prior work measuring skew in ad delivery, we develop a new methodology for black-box auditing of algorithms for discrimination in the delivery of job advertisements. Our first contribution is to identify the distinction between skew in ad delivery due to protected categories such as gender or race, from skew due to differences in qualification among people in the targeted audience. This distinction is important in U.S. law, where ads may be targeted based on qualifications, but not on protected categories. Second, we develop an auditing methodology that distinguishes between skew explainable by differences in qualifications from other factors, such as the ad platform’s optimization for engagement or training its algorithms on biased data. Our method controls for job qualification by comparing ad delivery of two concurrent ads for similar jobs, but for a pair of companies with different de facto gender distributions of employees. We describe the careful statistical tests that establish evidence of non-qualification skew in the results. Third, we apply our proposed methodology to two prominent targeted advertising platforms for job ads: Facebook and LinkedIn. We confirm skew by gender in ad delivery on Facebook, and show that it cannot be justified by differences in qualifications. We fail to find skew in ad delivery on LinkedIn. Finally, we suggest improvements to ad platform practices that could make external auditing of their algorithms in the public interest more feasible and accurate.

This paper was awarded runner-up for best student paper at The Web Conference 2021.

The data from this paper is upon request, please see our dataset page.

This work was reported in the popular press: The Intercept, MIT Technology Review, Wall Street Journal, The Register, VentureBeat, Reuters, The Verge, Engadget, Associated Press.

Categories
Presentations Publications

new poster “Measuring the Internet during Covid-19 to Evaluate Work-from-Home” at the NSF PREPARE-VO Workshop

Xiao Song presented the poster “Measuring the Internet during Covid-19 to Evaluate Work-from-Home (poster)” at the NSF PREPARE-VO Workshop on 2020-12-15. Xiao describes the poster in our video.

A case study network showing network changes as a result of work-from-home. Here we know ground truth and can see weekly work behavior (the groups of five bumps), followed by changes on the right in March when work-from-home begins.

There was no formal abstract, but this poster presents early results from examining Internet address changes to identify work-from-home resulting from Covid-19.

This work is part of the MINCEQ project, supported as an NSF CISE RAPID, NSF-2028279.

Categories
Publications Students

congratulations to Lan Wei for her new PhD

I would like to congratulate Dr. Lan Wei for defending her PhD in September 2020 and completing her doctoral dissertation “Anycast Stability, Security and Latency in The Domain Name System (DNS) and Content Deliver Networks (CDNs)” in December 2020.

From the abstract:

Clients’ performance is important for both Content-Delivery Networks (CDNs) and the Domain Name System (DNS). Operators would like the service to meet expectations of their users. CDNs providing stable connections will prevent users from experiencing downloading pause from connection breaks. Users expect DNS traffic to be secure without being intercepted or injected. Both CDN and DNS operators care about a short network latency, since users can become frustrated by slow replies.


Many CDNs and DNS services (such as the DNS root) use IP anycast to bring content closer to users. Anycast-based services announce the same IP address(es) from globally distributed sites. In an anycast infrastructure, Internet routing protocols will direct users to a nearby site naturally. The path between a user and an anycast site is formed on a hop-to-hop basis—at each hop} (a network device such as a router), routing protocols like Border Gateway Protocol (BGP) makes the decision about which next hop to go to. ISPs at each hop will impose their routing policies to influence BGP’s decisions. Without globally knowing (also unable to modify) the distributed information of BGP routing table of every ISP on the path, anycast infrastructure operators are unable to predict and control in real-time which specific site a user will visit and what the routing path will look like. Also, any change in routing policy along the path may change both the path and the site visited by a user. We refer to such minimal control over routing towards an anycast service, the uncertainty of anycast routing. Using anycast spares extra traffic management to map users to sites, but can operators provide a good anycast-based service without precise control over the routing?


This routing uncertainty raises three concerns: routing can change, breaking connections; uncertainty about global routing means spoofing can go undetected, and lack of knowledge of global routing can lead to suboptimal latency. In this thesis, we show how we confirm the stability, how we confirm the security, and how we improve the latency of anycast to answer these three concerns. First, routing changes can cause users to switch sites, and therefore break a stateful connection such as a TCP connection immediately. We study routing stability and demonstrate that connections in anycast infrastructure are rarely broken by routing instability. Of all vantage points (VPs), fewer than 0.15% VP’s TCP connections frequently break due to timeout in 5s during all 17 hours we observed. We only observe such frequent TCP connection break in 1 service out of all 12 anycast services studied. A second problem is DNS spoofing, where a third-party can intercept the DNS query and return a false answer. We examine DNS spoofing to study two aspects of security–integrity and privacy, and we design an algorithm to detect spoofing and distinguish different mechanisms to spoof anycast-based DNS. We show that DNS spoofing is uncommon, happening to only 1.7% of all VPs, although increasing over the years. Among all three ways to spoof DNS–injections, proxies, and third-party anycast site (prefix hijack), we show that third-party anycast site is the least popular one. Last, diagnosing poor latency and improving the latency can be difficult for CDNs. We develop a new approach, BAUP (bidirectional anycast unicast probing), which detects inefficient routing with better routing replacement provided. We use BAUP to study anycast latency. By applying BAUP and changing peering policies, a commercial CDN is able to significantly reduce latency, cutting median latency in half from 40ms to 16ms for regional users.

Lan defended her PhD when USC was on work-from-home due to COVID-19; she is the third ANT student with a fully on-line PhD defense.

Categories
Papers Publications

new journal paper “Plumb: Efficient Stream Processing of Multi-User Pipelines” in the Journal of Software: Practice and Experience

We have published a new journal paper “Plumb: Efficient Stream Processing of Multi-User Pipelines” in Wiley’s Journal of Software: Practice and Experience, available at https://onlinelibrary.wiley.com/doi/10.1002/spe.2909

Plumb provides a new pipeline-graph abstraction that allows multiple users to specify workflows in which Plumb can detect and elimiate duplicate processing and handle processing skew due to unbalanced data or stages. The end result is that users get their results faster and a shared cluster is efficiently utilized.

From the abstract of our journal paper:

Operational services run 24×7 and require analytics pipelines to evaluate performance. In mature services such as DNS, these pipelines often grow to many stages developed by multiple, loosely-coupled teams. Such pipelines pose two problems: first, computation and data storage may be duplicated across components developed by different groups, wasting resources. Second, processing can be skewed, with structural skew occurring when different pipeline stages need different amounts of resources, and computational skew occurring when a block of input data requires increased resources. Duplication and structural skew both decrease efficiency, increasing cost, latency, or both. Computational skew can cause pipeline failure or deadlock when resource consumption balloons; we have seen cases where pessimal traffic increases CPU requirements 6-fold. Detecting duplication is challenging when components from multiple teams evolve independently and require fault isolation. Skew management is hard due to dynamic workloads coupled with the conflicting goals of both minimizing latency and maximizing utilization. We propose Plumb, a framework to abstract stream processing as large-block streaming (LBS) for a multi-stage, multi-user workflow. Plumb users express analytics as a DAG of processing modules, allowing Plumb to integrate and optimize workflows from multiple users. Many real-world applications map to the LBS abstraction. Plumb detects and eliminates duplicate computation and storage, and it detects and addresses both structural and computational skew by tracking computation across the pipeline. We exercise Plumb using the analytics pipeline for B-Root DNS. We compare Plumb to a hand-tuned system, cutting latency to one-third the original, and requiring 39% fewer container hours, while supporting more flexible, multi-user analytics and providing greater robustness to DDoS-driven demands.

This journal paper is joint work of Abdul Qadeer and  John Heidemann from USC/ISI.

Plumb is open source software and we will be interested in beta testers. Please contact us if you think it would be useful to manage your workflows over one or a cluster of computers.

Categories
Papers Publications

new journal paper “Detecting IoT Devices in the Internet” in IEEE/ACM Transactions on Networking

We have published a new journal paper “Detecting IoT Devices in the Internet” in IEEE/ACM Transactions on Networking, available at https://www.isi.edu/~johnh/PAPERS/Guo20c.pdf

Figure 5 from [Guo20c] showing per-device-type AS penetrations from 2013 to 2018 for 16 of the 23 device types we studies (omitting 7 device types appearing in less than10 ASes)

From the abstract of our journal paper:

Distributed Denial-of-Service (DDoS) attacks launched from compromised Internet-of-Things (IoT) devices have shown how vulnerable the Internet is to largescale DDoS attacks. To understand the risks of these attacks requires learning about these IoT devices: where are they? how many are there? how are they changing? This paper describes three new methods to find IoT devices on the Internet: server IP addresses in traffic, server names in DNS queries, and manufacturer information in TLS certificates. Our primary methods (IP addresses and DNS names) use knowledge of servers run by the manufacturers of these devices. Our third method uses TLS certificates obtained by active scanning. We have applied our algorithms to a number of observations. With our IP-based algorithm, we report detections from a university campus over 4 months and from traffic transiting an IXP over 10 days. We apply our DNS-based algorithm to traffic from 8 root DNS servers from 2013 to 2018 to study AS-level IoT deployment. We find substantial growth (about 3.5×) in AS penetration for 23 types of IoT devices and modest increase in device type density for ASes detected with these device types (at most 2 device types in 80% of these ASes in 2018). DNS also shows substantial growth in IoT deployment in residential households from 2013 to 2017. Our certificate-based algorithm finds 254k IP cameras and network video recorders from 199 countries around the world.

We make operational traffic we captured from 10 IoT devices we own public at https://ant.isi.edu/datasets/iot/. We also use operational traffic of 21 IoT devices shared by University of New South Wales at http://149.171.189.1/.

This journal paper is joint work of Hang Guo and  John Heidemann from USC/ISI.

Categories
Publications Technical Report

new technical report: IoTSTEED: Bot-side Defense to IoT-based DDoS Attacks (Extended)

We have released a new technical report IoTSTEED: Bot-side Defense to IoT-based DDoS Attacks (Extended) as ISI-TR-738, available at https://www.isi.edu/~hangguo/papers/Guo20a.pdf.

From the abstract:

We show IoTSTEED runs
well on a commodity router: memory usage is small (4% of 512MB) and the router forwards traffic at full uplink rates despite about 50% of CPU usage.

We propose IoTSTEED, a system running in edge routers to defend against Distributed Denial-of-Service (DDoS) attacks launched from compromised Internet-of-Things (IoT) devices. IoTSTEED watches traffic that leaves and enters the home network, detecting IoT devices at home, learning the benign servers they talk to, and filtering their traffic to other servers as a potential DDoS attack. We validate IoTSTEED’s accuracy and false positives (FPs) at detecting devices, learning servers and filtering traffic with replay of 10 days of benign traffic captured from an IoT access network. We show IoTSTEED correctly detects all 14 IoT and 6 non-IoT devices in this network (100% accuracy) and maintains low false-positive rates when learning the servers IoT devices talk to (flagging 2% benign servers as suspicious) and filtering IoT traffic (dropping only 0.45% benign packets). We validate IoTSTEED’s true positives (TPs) and false negatives (FNs) in filtering attack traffic with replay of real-world DDoS traffic. Our experiments show IoTSTEED mitigates all typical attacks, regardless of the attacks’ traffic types, attacking devices and victims; an intelligent adversary can design to avoid detection in a few cases, but at the cost of a weaker attack. Lastly, we deploy IoTSTEED in NAT router of an IoT access network for 10 days, showing reasonable resource usage and verifying our testbed experiments for accuracy and learning in practice.

We share 10-day operational traffic captured from 14 IoT devices we own at https://ant.isi.edu/datasets/iot/ (see IoT_Operation_Traces-20200127) and release source code for IoTSTEED at https://ant.isi.edu/software/iotsteed/index.html.

This technical report is joint work of Hang Guo and John Heidemann from USC/ISI.

Categories
Papers Publications

New paper “Bidirectional Anycast/Unicast Probing (BAUP): Optimizing CDN Anycast” at IFIP TMA 2020

We published a new paper “Bidirectional Anycast/Unicast Probing (BAUP): Optimizing CDN Anycast” by Lan Wei (University of Southern California/ ISI), Marcel Flores (Verizon Digital Media Services), Harkeerat Bedi (Verizon Digital Media Services), John Heidemann (University of Southern California/ ISI) at Network Traffic Measurement and Analysis Conference 2020.

From the abstract:

IP anycast is widely used today in Content Delivery Networks (CDNs) and for Domain Name System (DNS) to provide efficient service to clients from multiple physical points-of-presence (PoPs). Anycast depends on BGP routing to map users to PoPs, so anycast efficiency depends on both the CDN operator and the routing policies of other ISPs. Detecting and diagnosing
inefficiency is challenging in this distributed environment. We propose Bidirectional Anycast/Unicast Probing (BAUP), a new approach that detects anycast routing problems by comparing anycast and unicast latencies. BAUP measures latency to help us identify problems experienced by clients, triggering traceroutes to localize the cause and suggest opportunities for improvement. Evaluating BAUP on a large, commercial CDN, we show that problems happens to 1.59% of observers, and we find multiple opportunities to improve service. Prompted by our work, the CDN changed peering policy and was able to significantly reduce latency, cutting median latency in half (40 ms to 16 ms) for regions with more than 100k users.

The data from this paper is publicly available from RIPE Atlas, please see paper reference for measurement IDs.

Categories
Papers

new paper “Improving Coverage of Internet Outage Detection in Sparse Blocks”

We will publish a new paper “Improving Coverage of Internet Outage Detection in Sparse Blocks” by Guillermo Baltra and John Heidemann in the Passive and Active Measurement Conference (PAM 2020) in Eugene, Oregon, USA, on March 30, 2020.

From the abstract:

There is a growing interest in carefully observing the reliability of the Internet’s edge. Outage information can inform our understanding of Internet reliability and planning, and it can help guide operations. Active outage detection methods provide results for more than 3M blocks, and passive methods more than 2M, but both are challenged by sparse blocks where few addresses respond or send traffic. We propose a new Full Block Scanning (FBS) algorithm to improve coverage for active scanning by providing reliable results for sparse blocks by gathering more information before making a decision. FBS identifies sparse blocks and takes additional time before making decisions about their outages, thereby addressing previous concerns about false outages while preserving strict limits on probe rates. We show that FBS can improve coverage by correcting 1.2M blocks that would otherwise be too sparse to correctly report, and potentially adding 1.7M additional blocks. FBS can be applied retroactively to existing datasets to improve prior coverage and accuracy.

This paper defines two algorithms: Full Block Scanning (FBS), to address false outages seen in active measurements of sparse blocks, and Lone Address Block Recovery (LABR), to handle blocks with one or two responsive addresses. We show that these algorithms increase coverage, from a nominal 67% (and as low as 53% after filtering) of responsive blocks before to 5.7M blocks, 96% of responsive blocks.
Categories
Publications Technical Report

new technical report “Peek Inside the Closed World: Evaluating Autoencoder-Based Detection of DDoS to Cloud ”

We have released a new technical report “Peek Inside the Closed World: Evaluating Autoencoder-Based Detection of DDoS to Cloud” as an ArXiv technical report 1912.05590, available at https://www.isi.edu/~hangguo/papers/Guo19a.pdf

We study 4 cloud IPs (SR1VP1 to 3 and SR2VP1) that are under attack. SR1VP3 sees a large number of mostly short DDoS events (71% of its 49 events being 1 second or less). SR1VP1 and SR1VP2 see smaller numbers of longer DDoS events (median duration for their 20 and 27 events are 121 and 140 seconds). SR2VP1 sees DDoS events of broad range of durations (from 1 second to more than 14 hours).

From the abstract of our technical report:

From the abstract:

Machine-learning-based anomaly detection (ML-based AD) has been successful at detecting DDoS events in the lab. However published evaluations of ML-based AD have only had limited data and have not provided insight into why it works. To address limited evaluation against real-world data, we apply autoencoder, an existing ML-AD model, to 57 DDoS attack events captured at 5 cloud IPs from a major cloud provider. To improve our understanding for why ML-based AD works or not works, we interpret this data with feature attribution and counterfactual explanation. We show that our version of autoencoders work well overall: our models capture nearly all malicious flows to 2 of the 4 cloud IPs under attacks (at least 99.99%) but generate a few false negatives (5% and 9%) for the remaining 2 IPs. We show that our models maintain near-zero false positives on benign flows to all 5 IPs. Our interpretation of results shows that our models identify almost all malicious flows with non-whitelisted (non-WL) destination ports (99.92%) by learning the full list of benign destination ports from training data (the normality). Interpretation shows that although our models learn incomplete normality for protocols and source ports, they still identify most malicious flows with non-WL protocols and blacklisted (BL) source ports (100.0% and 97.5%) but risk false positives. Interpretation also shows that our models only detect a few malicious flows with BL packet sizes (8.5%) by incorrectly inferring these BL sizes as normal based on incomplete normality learned. We find our models still detect a quarter of flows (24.7%) with abnormal payload contents even when they do not see payload by combining anomalies from multiple flow features. Lastly, we summarize the implications of what we learn on applying autoencoder-based AD in production.problme?Machine-learning-based anomaly detection (ML-based AD) has been successful at detecting DDoS events in the lab. However published evaluations of ML-based AD have only had limited data and have not provided insight into why it works. To address limited evaluation against real-world data, we apply autoencoder, an existing ML-AD model, to 57 DDoS attack events captured at 5 cloud IPs from a major cloud provider. To improve our understanding for why ML-based AD works or not works, we interpret this data with feature attribution and counterfactual explanation. We show that our version of autoencoders work well overall: our models capture nearly all malicious flows to 2 of the 4 cloud IPs under attacks (at least 99.99%) but generate a few false negatives (5% and 9%) for the remaining 2 IPs. We show that our models maintain near-zero false positives on benign flows to all 5 IPs. Our interpretation of results shows that our models identify almost all malicious flows with non-whitelisted (non-WL) destination ports (99.92%) by learning the full list of benign destination ports from training data (the normality). Interpretation shows that although our models learn incomplete normality for protocols and source ports, they still identify most malicious flows with non-WL protocols and blacklisted (BL) source ports (100.0% and 97.5%) but risk false positives. Interpretation also shows that our models only detect a few malicious flows with BL packet sizes (8.5%) by incorrectly inferring these BL sizes as normal based on incomplete normality learned. We find our models still detect a quarter of flows (24.7%) with abnormal payload contents even when they do not see payload by combining anomalies from multiple flow features. Lastly, we summarize the implications of what we learn on applying autoencoder-based AD in production.

This technical report is joint work of Hang Guo and John Heidemann from USC/ISI and Xun Fan, Anh Cao and Geoff Outhred from Microsoft

Categories
Papers Publications

new paper “Identifying Important Internet Outages” at the Sixth National Symposium for NSF REU Research in Data Science, Systems, and Security

We will publish a new paper “Identifying Important Internet Outages” by Ryan Bogutz, Yuri Pradkin, and John Heidemann, in the Sixth National Symposium for NSF REU Research in Data Science, Systems, and Security in Los Angeles, California, USA, on December 12, 2019.

From the abstract:

[Bogutz19a, figure 1]: Our sideboard showing important outages on 2019-03-08, including this outage in Venezuela.

Today, outage detection systems can track outages across the whole IPv4 Internet—millions of networks. However, it becomes difficult to find meaningful, interesting events in this huge dataset, since three months of data can easily include 660M observations and thousands of outage events. We propose an outage reporting system that sifts through this data to find the most interesting events. We explore multiple metrics to evaluate interesting”, reflecting the size and severity of outages. We show that defining interest as the product of size by severity works well, avoiding degenerate cases like complete outages affecting a few people, and apparently large outages that affect only a small fraction of people in an area. We have integrated outage reporting into our existing public website (https://outage.ant.isi.edu) with the goal of making near-real-time outage information accessible to the general public. Such data can help answer questions like “what are the most significant outages today?”, did Florida have major problems in an ongoing hurricane?”, and
“are there power outages in Venezuela?”.

The data from this paper is available publicly and in our website. The technical report ISI-TR-735 includes some additional data.