Stefan Schmid

Yuval Emek, Yuval Gil, Maciej Pacut et al.

We introduce a novel method for the rigorous quantitative evaluation of online algorithms that relaxes the "radical worst-case" perspective of classic competitive analysis. In contrast to prior work, our method, referred to as randomly infused advice (RIA), does not make any probabilistic assumptions about the input sequence and does not rely on the development of designated online algorithms. Rather, it can be applied to existing online randomized algorithms, introducing a means to evaluate their performance in scenarios that lie outside the radical worst-case regime. More concretely, an online algorithm ALG with RIA benefits from pieces of advice generated by an omniscient but not entirely reliable oracle. The crux of the new method is that the advice is provided to ALG by writing it into the buffer B from which ALG normally reads its random bits, hence allowing us to augment it through a very simple and non-intrusive interface. The (un)reliability of the oracle is captured via a parameter 0 {\le} α {\le} 1 that determines the probability (per round) that the advice is successfully infused by the oracle; if the advice is not infused, which occurs with probability 1 - α, then the buffer B contains fresh random bits (as in the classic online setting). The applicability of the new RIA method is demonstrated by applying it to three extensively studied online problems: paging, uniform metrical task systems, and online set cover. For these problems, we establish new upper bounds on the competitive ratio of classic online algorithms that improve as the infusion parameter α increases. These are complemented with (often tight) lower bounds on the competitive ratio of online algorithms with RIA for the three problems.

Julien Dallot, Caio Caldeira, Arash Pourdamghani et al.

This paper studies the integration of machine-learned advice in overlay networks in order to adapt their topology to the incoming demand. Such demand-aware systems have recently received much attention, for example in the context of data structures (Fu et al. in ICLR 2025, Zeynali et al. in ICML 2024). We in this paper extend this vision to overlay networks where requests are not to individual keys in a data structure but occur between communication pairs, and where algorithms have to be distributed. In this setting, we present an algorithm that adapts the topology (and the routing paths) of the overlay network to minimize the hop distance travelled by bit, that is, distance times demand. In a distributed manner, each node receives an (untrusted) prediction of the future demand to help him choose its set of neighbors and its forwarding table. This paper focuses on optimizing the well-known skip list networks (SLNs) for their simplicity. We start by introducing continuous skip list networks (C-SLNs) which are a generalization of SLNs specifically designed to tolerate predictive errors. We then present our learning-augmented algorithm, called LASLiN, and prove that its performance is (i) similar to the best possible SLN in case of good predictions ($O(1)$-consistency) and (ii) at most a logarithmic factor away from a standard overlay network in case of arbitrarily wrong predictions ($O(\log^2 n)$-robustness, where $n$ is the number of nodes in the network). Finally, we demonstrate the resilience of LASLiN against predictive errors (ie, its smoothness) using various error types on both synthetic and real demands.

Matthias Bentert, Shay Kutten, Darya Melnyk et al.

The theoretical model behind the pigeon post as a link layer in a communication network was introduced by Shannon (under the guise of studying One-Time Pads for cryptography). That is, to send a one-hop message to $v$, a node $u$ needs a mail pigeon bred and raised at $v$. When sending a message using a pigeon to $v$, node $u$ loses the pigeon. To send another message to $v$, node $u$ needs another pigeon of $v$. It has been demonstrated that the communication bandwidth achievable with pigeon post can exceed that of networks using other media. This has already motivated the introduction of Internet standards that allow the use of pigeons as Internet link-layer media. In this paper, we begin to fill in the missing piece: designing algorithms for breeding and scheduling pigeons to meet a given communication demand efficiently, minimizing the number of pigeons required. We consider singlehop, 2-hop, and multihop pigeon use. While the singlehop variant admits a simple characterization, both the 2-hop and the multihop variants are NP-hard. For the latter variants, we present a polynomial-time algorithm based on demand aggregation that achieves a 2-approximation for the number of pigeons used. We believe that this pigeon-based perspective offers both amusing and instructive insights into network design and hopefully, into ornithology.

Clemens Lottermoser, Simon Damm, Stefan Schmid

This paper presents an empirical study of dynamic factors affecting link quality in Low Earth Orbit (LEO) satellite communications, using Starlink as a case study. Over 56 days, 112 high-quality meteorological measurements in mostly 1-min intervals, co-located with a user terminal, were collected, alongside frequent network performance data. Cloud characteristics were estimated using professional weather instruments such as a ceilometer, microwave radiometer, and vision-language model on sky images. Our results show that general cloud presence does not significantly impact throughput or latency. The impact of cloud coverage rather depends on the presence of liquid water in the atmosphere, quantified by liquid water path (LWP), which correlates with notable download throughput reductions (up to 60 MBit/s), especially during rain. Upload and latency were largely unaffected. Analysis of the evolving satellite network revealed that newer satellite hardware and infrastructural upgrades also contributed to performance increases during the experiment period. These findings highlight atmospheric liquid water as the key weather-related factor affecting link quality and underscore the influence of network changes over time.

Daniel Grimm, Maximilian Zipfl, Felix Hertlein et al.

Precisely predicting the future trajectories of surrounding traffic participants is a crucial but challenging problem in autonomous driving, due to complex interactions between traffic agents, map context and traffic rules. Vector-based approaches have recently shown to achieve among the best performances on trajectory prediction benchmarks. These methods model simple interactions between traffic agents but don't distinguish between relation-type and attributes like their distance along the road. Furthermore, they represent lanes only by sequences of vectors representing center lines and ignore context information like lane dividers and other road elements. We present a novel approach for vector-based trajectory prediction that addresses these shortcomings by leveraging three crucial sources of information: First, we model interactions between traffic agents by a semantic scene graph, that accounts for the nature and important features of their relation. Second, we extract agent-centric image-based map features to model the local map context. Finally, we generate anchor paths to enforce the policy in multi-modal prediction to permitted trajectories only. Each of these three enhancements shows advantages over the baseline model HoliGraph.

Vamsi Addanki, Maciej Pacut, Stefan Schmid

Packet buffers in datacenter switches are shared across all the switch ports in order to improve the overall throughput. The trend of shrinking buffer sizes in datacenter switches makes buffer sharing extremely challenging and a critical performance issue. Literature suggests that push-out buffer sharing algorithms have significantly better performance guarantees compared to drop-tail algorithms. Unfortunately, switches are unable to benefit from these algorithms due to lack of support for push-out operations in hardware. Our key observation is that drop-tail buffers can emulate push-out buffers if the future packet arrivals are known ahead of time. This suggests that augmenting drop-tail algorithms with predictions about the future arrivals has the potential to significantly improve performance. This paper is the first research attempt in this direction. We propose Credence, a drop-tail buffer sharing algorithm augmented with machine-learned predictions. Credence can unlock the performance only attainable by push-out algorithms so far. Its performance hinges on the accuracy of predictions. Specifically, Credence achieves near-optimal performance of the best known push-out algorithm LQD (Longest Queue Drop) with perfect predictions, but gracefully degrades to the performance of the simplest drop-tail algorithm Complete Sharing when the prediction error gets arbitrarily worse. Our evaluations show that Credence improves throughput by $1.5$x compared to traditional approaches. In terms of flow completion times, we show that Credence improves upon the state-of-the-art approaches by up to $95\%$ using off-the-shelf machine learning techniques that are also practical in today's hardware. We believe this work opens several interesting future work opportunities both in systems and theory that we discuss at the end of this paper.

Ylli Sadikaj, Hongkuan Zhou, Lavdim Halilaj et al.

Precise optical inspection in industrial applications is crucial for minimizing scrap rates and reducing the associated costs. Besides merely detecting if a product is anomalous or not, it is crucial to know the distinct type of defect, such as a bent, cut, or scratch. The ability to recognize the "exact" defect type enables automated treatments of the anomalies in modern production lines. Current methods are limited to solely detecting whether a product is defective or not without providing any insights on the defect type, nevertheless detecting and identifying multiple defects. We propose MultiADS, a zero-shot learning approach, able to perform Multi-type Anomaly Detection and Segmentation. The architecture of MultiADS comprises CLIP and extra linear layers to align the visual- and textual representation in a joint feature space. To the best of our knowledge, our proposal, is the first approach to perform a multi-type anomaly segmentation task in zero-shot learning. Contrary to the other baselines, our approach i) generates specific anomaly masks for each distinct defect type, ii) learns to distinguish defect types, and iii) simultaneously identifies multiple defect types present in an anomalous product. Additionally, our approach outperforms zero/few-shot learning SoTA methods on image-level and pixel-level anomaly detection and segmentation tasks on five commonly used datasets: MVTec-AD, Visa, MPDD, MAD and Real-IAD.

Mélanie Cambus, Darya Melnyk, Tijana Milentijević et al.

Federated learning allows each client to keep its data locally when training machine learning models in a distributed setting. Significant recent research established the requirements that the input must satisfy in order to guarantee convergence of the training loop. This line of work uses averaging as the aggregation rule for the training models. In particular, we are interested in whether federated learning is robust to Byzantine behavior, and observe and investigate a tradeoff between the average/centroid and the validity conditions from distributed computing. We show that the various validity conditions alone do not guarantee a good approximation of the average. Furthermore, we show that reaching good approximation does not give good results in experimental settings due to possible Byzantine outliers. Our main contribution is the first lower bound of $\min\{\frac{n-t}{t},\sqrt{d}\}$ on the centroid approximation under box validity that is often considered in the literature, where $n$ is the number of clients, $t$ the upper bound on the number of Byzantine faults, and $d$ is the dimension of the machine learning model. We complement this lower bound by an upper bound of $2\min\{n,\sqrt{d}\}$, by providing a new analysis for the case $n<d$. In addition, we present a new algorithm that achieves a $\sqrt{2d}$-approximation under convex validity, which also proves that the existing lower bound in the literature is tight. We show that all presented bounds can also be achieved in the distributed peer-to-peer setting. We complement our analytical results with empirical evaluations in federated stochastic gradient descent and federated averaging settings.

Hongkuan Zhou, Lavdim Halilaj, Sebastian Monka et al.

Despite the remarkable success of deep neural networks (DNNs) in computer vision, they fail to remain high-performing when facing distribution shifts between training and testing data. In this paper, we propose Knowledge-Guided Visual representation learning (KGV) - a distribution-based learning approach leveraging multi-modal prior knowledge - to improve generalization under distribution shift. It integrates knowledge from two distinct modalities: 1) a knowledge graph (KG) with hierarchical and association relationships; and 2) generated synthetic images of visual elements semantically represented in the KG. The respective embeddings are generated from the given modalities in a common latent space, i.e., visual embeddings from original and synthetic images as well as knowledge graph embeddings (KGEs). These embeddings are aligned via a novel variant of translation-based KGE methods, where the node and relation embeddings of the KG are modeled as Gaussian distributions and translations, respectively. We claim that incorporating multi-model prior knowledge enables more regularized learning of image representations. Thus, the models are able to better generalize across different data distributions. We evaluate KGV on different image classification tasks with major or minor distribution shifts, namely road sign classification across datasets from Germany, China, and Russia, image classification with the mini-ImageNet dataset and its variants, as well as the DVM-CAR dataset. The results demonstrate that KGV consistently exhibits higher accuracy and data efficiency across all experiments.

Hongkuan Zhou, Lavdim Halilaj, Sebastian Monka et al.

Open-domain visual entity recognition aims to identify and link entities depicted in images to a vast and evolving set of real-world concepts, such as those found in Wikidata. Unlike conventional classification tasks with fixed label sets, it operates under open-set conditions, where most target entities are unseen during training and exhibit long-tail distributions. This makes the task inherently challenging due to limited supervision, high visual ambiguity, and the need for semantic disambiguation. We propose a Knowledge-guided Contrastive Learning (KnowCoL) framework that combines both images and text descriptions into a shared semantic space grounded by structured information from Wikidata. By abstracting visual and textual inputs to a conceptual level, the model leverages entity descriptions, type hierarchies, and relational context to support zero-shot entity recognition. We evaluate our approach on the OVEN benchmark, a large-scale open-domain visual recognition dataset with Wikidata IDs as the label space. Our experiments show that using visual, textual, and structured knowledge greatly improves accuracy, especially for rare and unseen entities. Our smallest model improves the accuracy on unseen entities by 10.5% compared to the state-of-the-art, despite being 35 times smaller.

Adrian Kuenzler, Stefan Schmid

Large language models (LLMs) are increasingly central to many applications, raising concerns about bias, fairness, and regulatory compliance. This paper reviews risks of biased outputs and their societal impact, focusing on frameworks like the EU's AI Act and the Digital Services Act. We argue that beyond constant regulation, stronger attention to competition and design governance is needed to ensure fair, trustworthy AI. This is a preprint of the Communications of the ACM article of the same title.

Philipp Zabka, Klaus-Tycho Foerster, Stefan Schmid et al.

Payment channel networks (PCNs) such as the Lightning Network offer an appealing solution to the scalability problem faced by many cryptocurrencies operating on a blockchain such as Bitcoin. However, PCNs also inherit the stringent dependability requirements of blockchain. In particular, in order to mitigate liquidity bottlenecks as well as on-path attacks, it is important that payment channel networks maintain a high degree of decentralization. Motivated by this requirement, we conduct an empirical centrality analysis of the popular Lightning Network, and in particular, the betweenness centrality distribution of the routing system. Based on our extensive data set (using several millions of channel update messages), we implemented a TimeMachine tool which enables us to study the network evolution over time. We find that although the network is generally fairly decentralized, a small number of nodes can attract a significant fraction of the transactions, introducing skew. Furthermore, our analysis suggests that over the last two years, the centrality has increased significantly, e.g., the inequality (measured by the Gini index) has increased by more than 10%.

Zeta Avarikioti, Krzysztof Pietrzak, Iosif Salem et al.

Payment channels effectively move the transaction load off-chain thereby successfully addressing the inherent scalability problem most cryptocurrencies face. A major drawback of payment channels is the need to ``top up'' funds on-chain when a channel is depleted. Rebalancing was proposed to alleviate this issue, where parties with depleting channels move their funds along a cycle to replenish their channels off-chain. Protocols for rebalancing so far either introduce local solutions or compromise privacy. In this work, we present an opt-in rebalancing protocol that is both private and globally optimal, meaning our protocol maximizes the total amount of rebalanced funds. We study rebalancing from the framework of linear programming. To obtain full privacy guarantees, we leverage multi-party computation in solving the linear program, which is executed by selected participants to maintain efficiency. Finally, we efficiently decompose the rebalancing solution into incentive-compatible cycles which conserve user balances when executed atomically. Keywords: Payment Channel Networks, Privacy and Rebalancing.

Krzysztof Pietrzak, Iosif Salem, Stefan Schmid et al.

Payment channel networks are a promising approach to improve the scalability of cryptocurrencies: they allow to perform transactions in a peer-to-peer fashion, along multi-hop routes in the network, without requiring consensus on the blockchain. However, during the discovery of cost-efficient routes for the transaction, critical information may be revealed about the transacting entities. This paper initiates the study of privacy-preserving route discovery mechanisms for payment channel networks. In particular, we present LightPIR, an approach which allows a source to efficiently discover a shortest path to its destination without revealing any information about the endpoints of the transaction. The two main observations which allow for an efficient solution in LightPIR are that: (1) surprisingly, hub labelling algorithms - which were developed to preprocess "street network like" graphs so one can later efficiently compute shortest paths - also work well for the graphs underlying payment channel networks, and that (2) hub labelling algorithms can be directly combined with private information retrieval. LightPIR relies on a simple hub labeling heuristic on top of existing hub labeling algorithms which leverages the specific topological features of cryptocurrency networks to further minimize storage and bandwidth overheads. In a case study considering the Lightning network, we show that our approach is an order of magnitude more efficient compared to a privacy-preserving baseline based on using private information retrieval on a database that stores all pairs shortest paths.

Sebastian Monka, Lavdim Halilaj, Stefan Schmid et al.

Traditional computer vision approaches, based on neural networks (NN), are typically trained on a large amount of image data. By minimizing the cross-entropy loss between a prediction and a given class label, the NN and its visual embedding space are learned to fulfill a given task. However, due to the sole dependence on the image data distribution of the training domain, these models tend to fail when applied to a target domain that differs from their source domain. To learn a more robust NN to domain shifts, we propose the knowledge graph neural network (KG-NN), a neuro-symbolic approach that supervises the training using image-data-invariant auxiliary knowledge. The auxiliary knowledge is first encoded in a knowledge graph with respective concepts and their relationships, which is then transformed into a dense vector representation via an embedding method. Using a contrastive loss function, KG-NN learns to adapt its visual embedding space and thus its weights according to the image-data invariant knowledge graph embedding space. We evaluate KG-NN on visual transfer learning tasks for classification using the mini-ImageNet dataset and its derivatives, as well as road sign recognition datasets from Germany and China. The results show that a visual model trained with a knowledge graph as a trainer outperforms a model trained with cross-entropy in all experiments, in particular when the domain gap increases. Besides better performance and stronger robustness to domain shifts, these KG-NN adapts to multiple datasets and classes without suffering heavily from catastrophic forgetting.

Lukas Aumayr, Esra Ceylan, Yannik Kopyciok et al.

Payment channel networks (PCNs) are among the most promising solutions to the scalability issues in permissionless blockchains, by allowing parties to pay each other off-chain through a path of payment channels (PCs). However, routing transactions comes at a cost which is proportional to the number of intermediaries, since each charges a fee for the routing service. Furthermore, analogous to other networks, malicious intermediaries in the payment path can lead to security and privacy threats. Virtual channels (VCs), i.e., bridges over PC paths, mitigate the above PCN issues, as an intermediary participates only once to set up the VC and is then excluded from every future VC transaction. However, similar to PCs, creating a VC has a cost that must be paid out of the bridged PCs' balance. Currently, we are missing guidelines to where and how many VCs to set up. Ideally, VCs should minimize transaction costs while mitigating security and privacy threats from on-path adversaries. In this work, we address for the first time the VC setup problem, formalizing it as an optimization problem. We present an integer linear program (ILP) to compute the globally optimal VC setup strategy in terms of transaction costs, security, and privacy. We then accompany the computationally heavy ILP with a fast local greedy algorithm. Our model and algorithms can be used with any on-path adversary, given that its strategy can be expressed as a set of corrupted nodes that is estimated by the honest nodes. We conduct an evaluation of the greedy algorithm over a snapshot of the Lightning Network (LN), the largest Bitcoin-based PCN. Our results confirm on real-world data that our greedy strategy minimizes costs while protecting against security and privacy threats of on-path adversaries. These findings may serve the LN community as guidelines for the deployment of VCs.

Apoorv Shukla, Kevin Hudemann, Zsolt Vági et al.

Is it possible to patch software bugs in P4 programs without human involvement? We show that this is partially possible in many cases due to advances in software testing and the structure of P4 programs. Our insight is that runtime verification can detect bugs, even those that are not detected at compile-time, with machine learning-guided fuzzing. This enables a more automated and real-time localization of bugs in P4 programs using software testing techniques like Tarantula. Once the bug in a P4 program is localized, the faulty code can be patched due to the programmable nature of P4. In addition, platform-dependent bugs can be detected. From P4_14 to P4_16 (latest version), our observation is that as the programmable blocks increase, the patchability of P4 programs increases accordingly. To this end, we design, develop, and evaluate P6 that (a) detects, (b) localizes, and (c) patches bugs in P4 programs with minimal human interaction. P6 tests P4 switch non-intrusively, i.e., requires no modification to the P4 program for detecting and localizing bugs. We used a P6 prototype to detect and patch seven existing bugs in eight publicly available P4 application programs deployed on two different switch platforms: behavioral model (bmv2) and Tofino. Our evaluation shows that P6 significantly outperforms bug detection baselines while generating fewer packets and patches bugs in P4 programs such as switch.p4 without triggering any regressions.

Kashyap Thimmaraju, Stefan Schmid

We revisit multi-tenant network virtualization in data centers, and make the case for tenant-specific virtual switches. In particular, tenant-specific virtual switches allow cloud providers to extend fine-grained billing (known, e.g., from serverless architectures) to the network, accounting not only for IO, but also CPU or energy. We sketch an architecture and present economical motivation and recent technological enablers. We also find that virtual switches today do not offer sufficient multi-tenancy and can introduce artificial performance bottlenecks, e.g., in load balancers. We conclude by discussing additional use cases for tentant-specific switches.

Utz Nisslmueller, Klaus-Tycho Foerster, Stefan Schmid et al.

Cryptocurrency off-chain networks such as Lightning (e.g., Bitcoin) or Raiden (e.g., Ethereum) aim to increase the scalability of traditional on-chain transactions. To support nodes in learning about possible paths to route their transactions, these networks need to provide gossip and probing mechanisms. This paper explores whether these mechanisms may be exploited to infer sensitive information about the flow of transactions, and eventually harm privacy. In particular, we identify two threats, related to an active and a passive adversary. The first is a probing attack: here the adversary aims to detect the maximum amount which is transferable in a given direction over a target channel by actively probing it and differentiating the response messages it receives. The second is a timing attack: the adversary discovers how close the destination of a routed payment actually is, by acting as a passive man-in-the middle and analyzing the time deltas between sent messages and their corresponding responses. We then analyze the limitations of these attacks and propose remediations for scenarios in which they are able to produce accurate results.

Saar Tochner, Stefan Schmid, Aviv Zohar

Off-chain transaction networks can mitigate the scalability issues of today's trustless electronic cash systems such as Bitcoin. However, these peer-to-peer networks also introduce a new attack surface which is not well-understood today. This paper identifies and analyzes, a novel Denial-of-Service attack which is based on route hijacking, i.e., which exploits the way transactions are routed and executed along the created channels of the network. This attack is conceptually interesting as even a limited attacker that manipulates the topology through the creation of new channels can navigate tradeoffs related to the way it attacks the network. Furthermore, the attack also highlights a fundamental design tradeoff for the defender (who determines its own routes): to become less predictable and hence secure, a rational node has to pay higher fees to nodes that forward its payments. We find that the three most common implementations for payment channels in Bitcoin (lnd, C-lightning, Eclair) approach routing differently. We begin by surveying the current state of the Lightning network and explore the routes chosen by these implementations. We find that in the current network nearly 60\% of all routes pass through only five nodes, while 80\% go through only 10 nodes. Thus, a relatively small number of colluding nodes can deny service to a large fraction of the network. We then turn to study an external attacker who creates links to the network and draws more routes through its nodes by asking for lower fees. We find that just five new links are enough to draw the majority (65\% - 75\%) of the traffic regardless of the implementation being used. The cost of creating these links is very low. We discuss the differences between implementations and eventually derive our own suggested routing policy, which is based on a novel combination of existing approaches.

Andrei-Alexandru Agape, Madalin Claudiu Danceanu, Rene Rydhof Hansen et al.

Emerging programmable dataplanes will revamp communication networks, allowing programmers to reconfigure and tailor switches towards their need, in a protocol-independent manner. While the community has articulated well the benefits of such network architectures in terms of flexibility and performance, little is known today about the security implications. We in this position paper argue that the programmable dataplanes in general and P4 in particular introduce an uncharted security landscape. In particular, we find that while some existing security studies on traditional OpenFlow-based networks still apply, P4 comes with several specific components and aspects which change the attack surface and introduce new challenges. We highlight several examples and provide a first systematic security analysis.

Kashyap Thimmaraju, Liron Schiff, Stefan Schmid

Today's routing protocols critically rely on the assumption that the underlying hardware is trusted. Given the increasing number of attacks on network devices, and recent reports on hardware backdoors this assumption has become questionable. Indeed, with the critical role computer networks play today, the contrast between our security assumptions and reality is problematic. This paper presents Software-Defined Adversarial Trajectory Sampling (SoftATS), an OpenFlow-based mechanism to efficiently monitor packet trajectories, also in the presence of non-cooperating or even adversarial switches or routers, e.g., containing hardware backdoors. Our approach is based on a secure, redundant and adaptive sample distribution scheme which allows us to provably detect adversarial switches or routers trying to reroute, mirror, drop, inject, or modify packets (i.e., header and/or payload). We evaluate the effectiveness of our approach in different adversarial settings, report on a proof-of-concept implementation, and provide a first evaluation of the performance overheads of such a scheme.

Liron Schiff, Stefan Schmid

Traffic inspection is a fundamental building block of many security solutions today. For example, to prevent the leakage or exfiltration of confidential insider information, as well as to block malicious traffic from entering the network, most enterprises today operate intrusion detection and prevention systems that inspect traffic. However, the state-of-the-art inspection systems do not reflect well the interests of the different involved autonomous roles. For example, employees in an enterprise, or a company outsourcing its network management to a specialized third party, may require that their traffic remains confidential, even from the system administrator. Moreover, the rules used by the intrusion detection system, or more generally the configuration of an online or offline anomaly detection engine, may be provided by a third party, e.g., a security research firm, and can hence constitute a critical business asset which should be kept confidential. Today, it is often believed that accounting for these additional requirements is impossible, as they contradict efficiency and effectiveness. We in this paper explore a novel approach, called Privacy Preserving Inspection (PRI), which provides a solution to this problem, by preserving privacy of traffic inspection and confidentiality of inspection rules and configurations, and e.g., also supports the flexible installation of additional Data Leak Prevention (DLP) rules specific to the company.