The Workshop on Analytics, Telemetry, and Cybersecurity for High-Performance Computing and Communications (HPCC) is newly launched and takes place in Taipei, Taiwan, on October 17, 2025, in conjunction with the ACM Conference on Computer and Communications Security (CCS'25). As its title suggests, the workshop centers on strengthening resilience and security in HPCC applications and infrastructures by leveraging leading technologies, including data-driven methodologies and machine intelligence techniques. The primary objective of this workshop is to provide a dedicated platform for researchers, practitioners, and industry experts to engage in discussions on cutting-edge topics in analytics, telemetry, and cybersecurity for HPCC. This year's call for contributions welcomes both full research papers and work-in-progress submissions, resulting in the acceptance of five full-length research papers. In addition, the workshop features a distinguished keynote presentation by Dr. Hsu-Chun Hsiao, Associate Professor in the Department of Computer Science and Information Engineering and the Graduate Institute of Networking and Multimedia at National Taiwan University. The WATCH'25 complete workshop proceedings can be found at: https://dl.acm.org/citation.cfm?id=3733826.

Deep neural network (DNN) models are susceptible to adversarial samples in white-box and opaque environments. Although previous studies have shown high attack success rates, coupling DNN models with interpretation models could offer a sense of security when a human expert is involved. However, in white-box environments, interpretable deep learning systems (IDLSes) have been shown to be vulnerable to malicious manipulations. As access to the components of IDLSes is limited in opaque settings, it becomes more challenging for the adversary to fool the system. In this work, we propose aQuery-efficientScore-based opaque attack against IDLSes, which requires no knowledge of the target model and its coupled interpretation model. By continuously refining the adversarial samples created based on feedback scores from the IDLS, our approach effectively reduces the number of model queries and navigates the search space to identify perturbations that can fool the system. We evaluate the attack's effectiveness on four convolutional neural network (CNN) models and two interpretation models, using both ImageNet and CIFAR datasets. Our results show that the proposed approach is query-efficient with a high attack success rate that can reach more than 95%, and an average transferability success rate of 69%. We have also demonstrated that our attack is resilient against various preprocessing defense techniques.

Deep learning methods have gained increasing attention in various applications due to their outstanding performance. For exploring how this high performance relates to the proper use of data artifacts and the accurate problem formulation of a given task, interpretation models have become a crucial component in developing deep learning-based systems. Interpretation models enable the understanding of the inner workings of deep learning models and offer a sense of security in detecting the misuse of artifacts in the input data. Similar to prediction models, interpretation models are also susceptible to adversarial inputs. This work introduces two attacks, AdvEdge and AdvEdge$^{+}$, which deceive both the target deep learning model and the coupled interpretation model. We assess the effectiveness of proposed attacks against four deep learning model architectures coupled with four interpretation models that represent different categories of interpretation models. Our experiments include the implementation of attacks using various attack frameworks. We also explore the attack resilience against three general defense mechanisms and potential countermeasures. Our analysis shows the effectiveness of our attacks in terms of deceiving the deep learning models and their interpreters, and highlights insights to improve and circumvent the attacks.

What a person chooses to share on social media can be very telling as to what their personality is. A person can show their personality through how many photos they post, what kinds of photos they post, and what kind of language they tend to use. This research investigates the correlation between a person's personality type and the photos and tweets which that person posts. Social media information and personality types were collected from two different platforms, SONA (college-age students) and MTurk (crowdsourcing platform). The same survey was given to both platforms. Participants answered questions pertaining to their personality, and based on these answers were assigned a True Color and scores for Big Five. Our results show that people who identified with the green True Color self schema tended to share more photos than the other True Color categories. Additionally, our results show that people with Open personality types tend to post more photos than other Big Five personality types, and those photos tend to have more people in them than other personality types.

Website fingerprinting (WF) allows a passive local eavesdropper to monitor the encrypted channel where users search the Internet and determine which website the user is visiting from the recorded traffic. The effectiveness of using deep learning (DL) in WF attacks has been explored in recent work. However, they all are built and evaluated on one-page traces. Our goal is to explore whether deep learning can be used to handle the situations when the captured traces are not best-case for an adversary, such as partial traces and two-page traces. We aim to reduce the distance between the lab experiments and the realistic conditions. We evaluate our proposed method in both closed-world and open-world settings and found that Convolutional Neural Network (CNN) outperforms Long-Short Term Memory network (LSTM) in all scenarios. CNN also shows a great potential in predicting on a smaller number of packets. For partial trace missing 20% packets in the beginning of the trace, the accuracy is improved from 8.28% to 86.93% compared to the original DL model by adding the head detection. We then show the accuracy of predicting on two-page traces. With an overlap of 80% between two websites, we are able to achieve an accuracy of 89.25% and 74.2% for the first and second website in the closed-world evaluation, and 95.5% and 75% in the open world from our simulation. To verify our simulation results, we set up a crawler to collect both training and testing data and gathered the largest two-page traces testing dataset ever used. The results shown in the real world experiment is consistent with the simulation.

Alternating direction method of multipliers (ADMM) is a widely used tool for machine learning in distributed settings where a machine learning model is trained over distributed data sources through an interactive process of local computation and message passing. Such an iterative process could cause privacy concerns of data owners. The goal of this paper is to provide differential privacy for ADMM-based distributed machine learning. Prior approaches on differentially private ADMM exhibit low utility under high privacy guarantee and assume the objective functions of the learning problems to be smooth and strongly convex. To address these concerns, we propose a novel differentially private ADMM-based distributed learning algorithm called DP-ADMM, which combines an approximate augmented Lagrangian function with time-varying Gaussian noise addition in the iterative process to achieve higher utility for general objective functions under the same differential privacy guarantee. We also apply the moments accountant method to analyze the end-to-end privacy loss. The theoretical analysis shows that the DP-ADMM can be applied to a wider class of distributed learning problems, is provably convergent, and offers an explicit utility-privacy tradeoff. To our knowledge, this is the first paper to provide explicit convergence and utility properties for differentially private ADMM-based distributed learning algorithms. The evaluation results demonstrate that our approach can achieve good convergence and model accuracy under high end-to-end differential privacy guarantee.

Most privacy-conscious users utilize HTTPS and an anonymity network such as Tor to mask source and destination IP addresses. It has been shown that encrypted and anonymized network traffic traces can still leak information through a type of attack called a website fingerprinting (WF) attack. The adversary records the network traffic and is only able to observe the number of incoming and outgoing messages, the size of each message, and the time difference between messages. In previous work, the effectiveness of website fingerprinting has been shown to have an accuracy of over 90% when using Tor as the anonymity network. Thus, an Internet Service Provider can successfully identify the websites its users are visiting. One main concern about website fingerprinting is its practicality.

Website fingerprinting attacks attempt to identify the website visited in anonymized and encrypted network traffic, that is, even if a user is using Tor and HTTPS. These attacks have been shown to be effective. Mitigations have been proposed which decreased the accuracy of the attacks from about 90% to about 20%. We propose a new mitigation technique based on traffic morphing and clustering. The intuition is that a lot of websites, by nature, are similar and can be clustered together. It is then easier and more efficient to make that whole cluster look exactly the same by using traffic morphing, rather than adding noise to make all websites look similar. All the websites in a cluster, thus, would become indistinguishable. There are many ways to perform traffic morphing. As a proof of concept, we used biggest, which means that all websites in a cluster will look exactly like the biggest website (in terms of network packet size) of that cluster. In simulating our proposed approach, the fingerprinting accuracy dropped from 70% to less than 1%.

Website fingerprinting attacks have been shown to be able to predict the website visited even if the network connection is encrypted and anonymized. These attacks have achieved accuracies as high as 92%. Mitigations to these attacks are using cover/decoy network traffic to add noise, padding to ensure all the network packets are the same size, and introducing network delays to confuse an adversary. Although these mitigations have been shown to be effective, reducing the accuracy to 10%, the overhead is very high. The latency overhead is above 100% and the bandwidth overhead is at least 40%. We introduce a new realistic cover traffic algorithm, based on a user's previous network traffic, to mitigate website fingerprinting attacks. In simulations, our algorithm reduces the accuracy of attacks to 14% with zero latency overhead and about 20% bandwidth overhead.

Many people now own smartphones and store all their documents such as pictures and financial statements on their phone. To protect this sensitive information, people generally use a passcode to prevent unauthorised access to their phone. Shoulder‐surfing attacks are well known. However, contrary to common belief, they are not easy to carry out. Shoulder‐surfing attacks to predict the passcode by humans are shown to not be accurate. The authors thus propose an automated algorithm to accurately predict the passcode entered by a victim on her smartphone by recording the video. Their proposed algorithm is able to predict over 92% of numbers entered in fewer than 75 s with training performed once.

Embedded healthcare devices are becoming increasingly popular. A device could use a low-power sensor to send data. These low-power sensors typically use ZigBee wireless signals; however, these signals are easily masked by stronger WiFi signals. If a sensor is sending data at the same time as a WiFi-enabled device, the sensor data is lost and needs to be retransmitted. This causes a loss in efficiency and energy usage of the sensor. We proposed two algorithms, based on a Hidden Markov Model and concordance, that are able to predict when a WiFi interference will occur, which ensures with high accuracy that the ZigBee sensor can send data without any risk of wireless interference. Concordance is based on the premise that past history repeats itself. Our experimental results are based on real datasets. Our Hidden Markov Model has a prediction accuracy rate of over 78% and our concordance algorithm has a prediction accuracy rate of over 70%. Our algorithms can also predict longer periods of time and achieve a higher prediction accuracy than current algorithms. Our concordance algorithm has a very low processing overhead, taking microseconds to make a prediction. The throughput improvement gain of our proposed algorithms over current algorithms is a factor of 2.5.

The threat of mass surveillance and the need for privacy have become mainstream recently. Most of the anonymity schemes have focused on Internet privacy. We propose an anonymity scheme for cellular phone calls. The cellular phones form an ad-hoc network relaying phone conversations through direct wifi connections. A proof-of-concept implementation on an Android smartphone is completed and shown to work with minimal delay in communications.

Vuze is a popular file‐sharing client. When looking for content, Vuze selects from its list of neighbours, a set of 20 nodes to be contacted; the selection is performed such that the neighbours closest to the content in terms of Vuze ID are contacted first. To improve efficiency of its searches, Vuze implements a network coordinate system: from the set of 20 to‐be‐contacted nodes, queries are sent to the closest nodes in terms of network distance, which is calculated by the difference in network coordinates. However, network coordinate systems are inherently insecure and a malicious peer can lie about its coordinate to appear closest to every peer in the network. This allows the malicious peer to bias next‐hop choices for victim peers such that queries will be sent to the attacker, thus hijacking every search query. In our experiments, almost 20% of the search queries are hijacked; the cost of performing this attack is minimal – less than $112/month.

We performed two surveys to understand how members of a university managed their passwords. At password creation, the university offered people four pre-generated random passwords, with the option of creating their own subject to stringent requirements. All passwords expired after 120days. We found that most respondents chose to create their own password and utilized coping strategies that undermined the security of the requirements, as well as reporting that the expiration times were too short. We also attempt to connect these behaviors to respondents' other password habits and demographics. We conclude that pre-generated random passwords, stringent password requirements, and rapid password expiration dates are unusable security requirements for most people and lead users to subvert password requirements and reuse passwords. Copyright © 2015 John Wiley & Sons, Ltd.

Abstract This paper describes an improved authentication method for mobile devices compared with commonly used authentication using characters (text passwords) and Personal Identification Number (PINs). The method introduced in this paper was developed to overcome the difficulty of remembering unfamiliar secrets such as a complex sequence of text characters while maintaining authentication strength. We introduce a new authentication approach, SmartPass, that works for both legacy systems and mobile devices. SmartPass is a web‐based mobile interface that provides a method to use visual map location secrets, which are theorized to be easier to remember and use than passwords and remain difficult for an adversary to guess. We evaluate the security strength of such a map‐based secret compared with traditional text password, PIN, and pattern secrets. This paper also seeks to evaluate user feedback as to the usability and memorability of this authentication method. Copyright © 2015 John Wiley & Sons, Ltd.

Tor is a popular anonymity-providing network used by over 500,000 users daily. The Tor network is made up of volunteer relays. To anonymously connect to a server, a user first creates a circuit, consisting of three relays, and routes traffic through these proxies before connecting to the server. The client is thus hidden from the server through three Tor proxies. If the three Tor proxies used by the client could be identified, the anonymity of the client would be reduced. One particular way of identifying the three Tor relays in a circuit is to perform a circuit clogging attack. This attack requires the client to connect to a malicious server (malicious content, such as an advertising frame, can be hosted on a popular server). The malicious server alternates between sending bursts of data and sending little traffic. During the burst period, the three relays used in the circuit will take longer to relay traffic due to the increase in processing time for the extra messages. If Tor relays are continuously monitored through network latency probes, an increase in network latency indicates that this Tor relay is likely being used in that circuit. We show, through experiments on the real Tor network, that the Tor relays in a circuit can be identified. A detection scheme is also proposed for clients to determine whether a circuit clogging attack is happening. The costs for both the attack and the detection mechanism are small and feasible in the current Tor network.

Cloud computing is an emerging field in computer science. Users are utilizing less of their own existing resources, while increasing usage of cloud resources. With the emergence of new technologies such as mobile devices, these devices are usually under-utilized, and can provide similar functionality to a cloud provided they are properly configured and managed. This paper proposes a Distributed Cloud Architecture to make use of independent resources provided by the devices/users. Resource discovery and allocation is critical in designing an efficient and practical distributed cloud. We propose using multi-valued distributed hash tables for efficient resource discovery. Leveraging the fact that there are many users providing resources such as CPU and memory, we define these resources under one key to easily locate devices with equivalent resources. We then propose a new auction mechanism, using a reserve bid formulated rationally by each user for the optimal allocation of discovered resources. Then we discuss how the Nash Equilibrium is achieved based on user requirements.

Abstract The Kad network, an implementation of the Kademlia DHT protocol, supports the popular eDonkey peer‐to‐peer file sharing network and has over 1 million concurrent nodes. We describe several attacks that exploit critical design weaknesses in Kad to allow an attacker with modest resources to cause a significant fraction of all searches to fail. We measure the cost and effectiveness of these attacks against a set of 16 000 nodes connected to the operational Kad network. Using our large‐scale simulator, DVN, we successfully scaled up to a 200 000 node experiment. We also measure the cost of previously proposed, generic DHT attacks against the Kad network and find that our attacks are much more cost effective. Finally, we introduce and evaluate simple mechanisms to significantly increase the cost of these attacks. Copyright © 2010 John Wiley & Sons, Ltd.

A network coordinate system [7, 14, 15] assigns virtual coordinates (network positions) to every node in the network. These coordinates are assigned so that the coordinate distance between two nodes reflects the real network distance between those two nodes. This allows any peer in the sytem to accurately estimate the network distance between any pair of nodes, without having the pair of nodes contact each other. Network coordinate systems' ability to predict the network latency between arbitrary pairs of nodes can be used in many applications: finding the closest node to download content from in a content distribution network or route to in a peer-to-peer system [18], reducing inter-ISP communication [5, 13], reducing the amount of state stored in routers [1], performing byzantine leader elections [6], and detecting Sybil attackers [3, 8].

A network coordinate system assigns Euclidean “virtual” coordinates to every node in a network to allow easy estimation of network latency between pairs of nodes that have never contacted each other. These systems have been implemented in a variety of applications, most notably the popular Vuze BitTorrent client. Zage and Nita-Rotaru (at CCS 2007) and independently, Kaafar et al. (at SIGCOMM 2007), demonstrated that several widely-cited network coordinate systems are prone to simple attacks, and proposed mechanisms to defeat these attacks using outlier detection to filter out adversarial inputs. Kaafar et al. goes a step further and requires that a fraction of the network is trusted. More recently, Sherr et al. (at USENIX ATC 2009) proposed Veracity, a distributed reputation system to secure network coordinate systems. We describe a new attack on network coordinate systems, Frog-Boiling, that defeats all of these defenses. Thus, even a system with trusted entities is still vulnerable to attacks. Moreover, having witnesses vouch for your coordinates as in Veracity does not prevent our attack. Finally, we demonstrate empirically that the Frog-Boiling attack is more disruptive than the previously known attacks: systems that attempt to reject “bad” inputs by statistical means or reputation cannot be used to secure a network coordinate system.

A network latency estimation scheme associates a "position" to every peer in a distributed network such that the latency between any two nodes can be accurately estimated from their positions. Applications for these schemes include efficient overlay construction, compact routing, anonymous route selection, and efficient byzantine agreement. We present a new latency estimation scheme, Treeple. Our scheme is different from existing ones in several aspects: Treeple is provably secure, rather than being able to resist known attacks; positions in Treeple are not Euclidean coordinates and reflect the underlying network topology; finally, positions in Treeple are accurate, stable, and can be assigned to peers not participating in the system.

Low-latency anonymity systems such as Tor, AN.ON, Crowds, and Anonymizer.com aim to provide anonymous connections that are both untraceable by “local” adversaries who control only a few machines and have low enough delay to support anonymous use of network services like Web browsing and remote login. One consequence of these goals is that these services leak some information about the network latency between the sender and one or more nodes in the system. We present two attacks on low-latency anonymity schemes using this information. The first attack allows a pair of colluding Web sites to predict, based on local timing information and with no additional resources, whether two connections from the same Tor exit node are using the same circuit with high confidence. The second attack requires more resources but allows a malicious Web site to gain several bits of information about a client each time he visits the site. We evaluate both attacks against two low-latency anonymity protocols—the Tor network and the MultiProxy proxy aggregator service—and conclude that both are highly vulnerable to these attacks.

Modern advanced botnets may employ a decentralized peer-to-peer overlay network to bootstrap and maintain their command and control channels, making them more resilient to traditional mitigation efforts such as server incapacitation. As an alternative strategy, the malware defense community has been trying to identify the bot-infected hosts and enumerate the IP addresses of the participating nodes so that the list can be used by system administrators to identify local infections, block spam emails sent from bots, and configure firewalls to protect local users. Enumerating the infected hosts, however, has presented challenges. One cannot identify infected hosts behind firewalls or NAT devices by employing crawlers, a commonly used enumeration technique where recursive get-peerlist lookup requests are sent newly discovered IP addresses of infected hosts. As many bot-infected machines in homes or offices are behind firewall or NAT devices, these crawler-based enumeration methods would miss a large portions of botnet infections. In this paper, we present the Passive P2P Monitor (PPM), which can enumerate the infected hosts regardless whether or not they are behind a firewall or NAT. As an empirical study, we examined the Storm botnet and enumerated its infected hosts using the PPM. We also improve our PPM design by incorporating a FireWall Checker (FWC) to identify nodes behind a firewall. Our experiment with the peer-to-peer Storm botnet shows that more than 40% of bots that contact the PPM are behind firewall or NAT devices, implying that crawler-based enumeration techniques would miss out a significant portion of the botnet population. Finally, we show that the PPM's coverage is based on a probability-based coverage model that we derived from the empirical observation of the Storm botnet.

A Distributed Hash Table (DHT) is a structured overlay network service that provides a decentralized lookup for mapping objects to locations. In this paper, we study the lookup performance of locating nodes responsible for replicated information in Kad - one of the largest DHT networks existing currently. Throughout the measurement study, we found that Kad lookups locate only 18% of nodes storing replicated data. This failure leads to limited reliability and an inefficient use of resources during lookups. Ironically, we found that this poor performance is due to the high level of routing table similarity, despite the relatively high churn rate in the network. We propose solutions which either exploit the high routing table similarity or avoid the duplicate returns using multiple target keys.

The Kad network, an implementation of the Kademlia DHT protocol, supports the popular eDonkey peer-to-peer file sharing network and has over 1 million concurrent nodes. We describe several attacks that exploit critical design weaknesses in Kad to allow an attacker with modest resources to cause a significant fraction of all searches to fail. We measure the cost and effectiveness of these attacks against a set of 16,000 nodes connected to the operational Kad network. We also measure the cost of previously proposed, generic DHT attacks against the Kad network and find that our attacks are much more cost effective. Finally, we introduce and evaluate simple mechanisms to significantly increase the cost of these attacks.

Low-latency anonymity systems such as Tor, AN.ON, Crowds, and Anonymizer.com aim to provide anonymous connections that are both untraceable by "local" adversaries who control only a few machines, and have low enough delay to support anonymous use of network services like web browsing and remote login. One consequence of these goals is that these services leak some information about the network latency between the sender and one or more nodes in the system. This paper reports on three experiments that partially measure the extent to which such leakage can compromise anonymity. First, using a public dataset of pairwise round-trip times (RTTs) between 2000 Internet hosts, we estimate that on average, knowing the network location of host A and the RTT to host B leaks 3.64 bits of information about the network location of B. Second, we describe an attack that allows a pair of colluding web sites to predict, based on local timing information and with no additional resources, whether two connections from the same Tor exit node are using the same circuit with 17% equal error rate. Finally, we describe an attack that allows a malicious website, with access to a network coordinate system and one corrupted Tor router, to recover roughly 6.8 bits of network location per hour.