To combat the rapid spread of misinformation on social networks, automated misinformation detection systems based on deep neural networks (DNNs) have been developed. However, these tools are often proprietary and lack transparency, which limits their usefulness. Furthermore, privacy concerns limit data sharing by data owners as well as by data-driven misinformation-detection services. Although data encryption techniques can help address privacy concerns in DNN inference, there is a challenge to the seamless integration of these techniques due to the encryption errors induced by cascaded encrypted operations, as well as a mismatch between the tools used for DNNs and cryptography. In this paper, we make two-fold contributions. First, we study the noise bounds of homomorphic encryption (HE) operations as error propagation in DNN layers and derive two properties that, if satisfied by the layer, will considerably reduce the output error. We identify that$L_{2}$regularization and sigmoid activation satisfy these properties and validate our hypothesis, for instance, replacing ReLU with sigmoid reduced the output error by$10^{6}\times$(best case) to$10\times$(worst case). Second, we extend the Python encryption library TenSeal by enabling the automatic conversion of a TensorFlow DNN into an encryption-compatible DNN with a few lines of code. These contributions are significant as encryption-friendly DL architectures are sorely needed to close the gap between DL-in-research and DL-in-practice.

Deep Learning (DL) algorithms have shown wonders in many Natural Language Processing (NLP) tasks such as language-to-language translation, spam filtering, fake-news detection, and comprehension understanding. However, research has shown that the adversarial vulnerabilities of deep learning networks manifest themselves when DL is used for NLP tasks. Most mitigation techniques proposed to date are supervised—relying on adversarial retraining to improve the robustness—which is impractical. This work introduces a novel, unsupervised detection methodology for detecting adversarial inputs to NLP classifiers. In summary, we note that minimally perturbing an input to change a model’s output—a major strength of adversarial attacks—is a weakness that leaves unique statistical marks reflected in the cumulative contribution scores of the input. Particularly, we show that the cumulative contribution score, called CF-score, of adversarial inputs is generally greater than that of the clean inputs. We thus propose Con-Detect—a Contribution based Detection method—for detecting adversarial attacks against NLP classifiers. Con-Detect can be deployed with any classifier without having to retrain it. We experiment with multiple attackers—Text-bugger, Text-fooler, PWWS—on several architectures—MLP, CNN, LSTM, Hybrid CNN-RNN, BERT—trained for different classification tasks—IMDB sentiment classification, fake-news classification, AG news topic classification—under different threat models—Con-Detect-blind attacks, Con-Detect-aware attacks, and Con-Detect-adaptive attacks—and show that Con-Detect can reduce the attack success rate (ASR) of different attacks from 100% to as low as 0% for the best cases and ≈70% for the worst case. Even in the worst case, we note a 100% increase in the required number of queries and a 50% increase in the number of words perturbed, suggesting that Con-Detect is hard to evade.

In this paper, we investigate the soft-reconfiguration of optical wireless data centers (WDCs). In the considered physical topology, edge top-of-rack (ToR) switches in the leaf layer are inter-connected with core switches in the spine layer via wavelength division multiplexing (WDM) based free-space optical (FSO) links. We propose an agile load balancing (LB) solution, namely SoftFG, to cope with the dynamically changing link load variations and the low-utilization time intervals within the wireless data centers (DCs). SoftFG executes flow grooming (FG) and soft reconfigurations on the virtual topology depending upon the fine-grain network statistics. Unlike the long-term LBs, SoftFG offloads large flows of congested paths onto underutilized links without making any hardware reconfiguration on path capacity and routes. Flows can be offloaded to other wavelengths within the same FSO link (i.e., intra-link), to other FSO links (i.e., inter-link), or within/across topologies (i.e., intra/inter topology). To do so, SoftFG ensures clear visibility on network paths, early congestion detection, and fast-accurate reaction to reroute offloaded flows onto underutilized wavelengths or links. Therefore, SoftFG is designed as a kernel module installed on the virtual switches/hypervisor. The module collects flow statistics based on a source-destination collaborative scheme and records them in flow and path information tables. SoftFG accordingly makes quick decisions on offloading and reroutes flows with high accuracy. Emulation results show that SoftFG delivers about 12 and 17 faster flow completion time (FCT) than LetFlow and CONGA LBs, respectively.

LightFDG is an integrated approach to flow detection (FD) and flow grooming (FG) in optical wireless data center networks (DCNs), which is interconnected via wavelength division multiplexing (WDM) based free-space optical (FSO) links. Since forwarding bandwidth-hungry elephant flows (EFs) and delay-sensitive mice flows (MFs) on the same path can cause severe performance degradation, the LightFDG optically grooms flows of each class into rack-to-rack (R2R) flows. Then, R2R-MF and R2R-EF flows are separately forwarded over lightpaths of separate MF and EF virtual topologies, respectively. Lightpaths are provisioned by jointly determining the capacity and route based on flows' arrival rate, size, and completion time request. To prevent EFs from congesting the MF lightpaths, high speed and accurate flow-detection mechanisms are also necessary for classifying EFs as soon as possible. Therefore, a fast-lightweight-and-accurate flow detection framework is developed by leveraging the transmission control protocol (TCP) behaviors. The proposed FD scheme has the flexibility of being implemented as in-network or centralized to classify flows of modifiable and unmodifiable hosts, respectively. Since the centralized scheme incurs considerable overhead, the processing and communication overhead is also mitigated by proposed techniques. Numerical results show that LightFDG outperforms traditional load balancers by about 3× for EFs and 10× for MFs. Along with the developed overhead mitigation methods, the centralized scheme is shown to provide up to 62× lower overhead with 100% accuracy and with about 224× higher detection speeds than the existing centralized solutions.

Traditional wired data center networks (DCNs) suffer from cabling complexity, lack flexibility, and are limited by the speed of digital switches. In this paper, we alternatively develop a top-down traffic grooming (TG) approach to the design and provisioning of mission-critical optical wireless DCNs. While switches are modeled as hybrid optoelectronic cross-connects, links are modeled as wavelength division multiplexing capable free-space optic channels. Using the standard TG terminology, we formulate the optimal mixed-integer TG problem considering the virtual topology, flow conversation, connection topology, non-bifurcation, and capacity constraints. Thereafter, we develop a fast yet efficient sub-optimal solution, which grooms mice flows (MFs), mission-critical flows (CFs), and forward on predetermined rack-to-rack (R2R) lightpaths. On the other hand, elephant flows (EFs) are forwarded over dedicated server-to-server express lightpaths whose routes and capacity are dynamically determined based on the availability of wavelength and capacity. To prioritize the CFs, we consider low and high-priority queues and analyze the delay characteristics such as waiting times, maximum hop counts, and blocking probability. As a result of grooming, the sub-wavelength traffic and adjusting the wavelength capacities, numerical results show that the proposed solutions can achieve significant performance enhancement by utilizing the bandwidth more efficiently, completing the flows faster than delay sensitivity requirements, and avoiding the traffic congestion by treating EFs and MFs separately.

Wireless data centers (DCs) are enablers of reconfigurable data center network (DCN) topologies by augmenting the cabling complexity and inflexibility of traditional wired DCs. In this paper, we propose an optical traffic grooming (TG) for mice flows (MFs) and elephant flows (EFs) in a wireless DCN which is interconnected with free-space optical (FSO) links operating on wavelength division multiplexing (WDM). Since handling the bandwidth-hungry EFs along with delay-sensitive MFs over the same network resources have undesirable consequences, proposed TG policy treat MFs and EFs separately. MFs/EFs destined to the same rack are groomed into larger rack-to-rack MF/EF flows over dedicated lightpaths whose routes and capacities are jointly determined taking the load balancing into account. Performance evaluations of proposed TG policy show a significant throughput improvement thanks to bandwidth efficient utilization of the wireless links. Therefore, proposed TG requires expeditious flow detection mechanisms which can immediately classify EFs with very high accuracy. Since these demands cannot be met by existing sampling and port-mirroring based solutions, we propose a lightweight and fast in-network flow detection (LightFD) mechanism. LightFD is designed as a module on the Virtual-Switch/Hypervisor, which detects EFs based on acknowledgment sequence number of flow packets. Emulation results show that LightFD can provide up to 110 times faster detection speeds than sampling-based methods with %100 detection accuracy. We also demonstrate that the EF detection speed has a considerable impact on achievable EF throughput.

In this work, we propose FLight, a fast, lightweight and adaptive mechanism for detecting elephant-flows while improving the detection accuracy and speed. FLight leverages the TCP communication behavior for its detection algorithm, it demonstrates a 100% elephant-flow detection accuracy, and is 242× faster than other centralized solutions.

Software Defined Networking (SDN) provides network engineers with a high-level of abstraction to manage network traffic and control the associated network resources. Unfortunately, the data-plane devices communicate with the controller for every new flow, which adds an extra overhead and causes excessive delays. Such communication relies on the probability called matching probability. In this work, we propose a mathematical model for the SDN flow-setup process with the consideration of all factors that contribute into the matching probability, such as proactive/reactive flow setup modes. Finally, we attempt at deriving the overall system capacity and blocking probability.

Traditional wired data center networks (DCNs) suffer from cabling complexity, lack flexibility, and are limited by the speed of digital switches. In this paper, we alternatively develop a top-down traffic grooming (TG) approach for design and provisioning of optical wireless DCNs. While switches are modeled as hybrid opto-electronic cross-connects, links are modeled as wavelength division multiplexing (WDM) capable free-space optic (FSO) channels. Using the standard TG terminology, we formulate the optimal mixed integer linear problem considering the virtual topology, flow conversation, connection topology, non-bifurcation, and capacity constraints. Thereafter, we develop a fast sub-optimal solution where mice flows (MFs) are groomed and forwarded on predetermined rack-to-rack (R2R) lightpaths. On the other hand, elephant flows (EFs) are forwarded over dedicated server-to-server (S2S) express lightpaths whose routes and capacity are dynamically determined based on wavelength and capacity availability. Emulation results show that proposed models and algorithms provide a significant throughput improvement upon traditional DCNs for both MFs and EFs.

The controller is a critical piece of the SDN architecture, where it is considered as the mastermind of SDN networks. Thus, its failure will cause a significant portion of the network to fail. Overload is one of the common causes of failure since the controller is frequently invoked by new flows. Even through SDN controllers are often replicated, the significant recovery time can be an overkill for the availability of the entire network. In order to overcome the problem of the overloaded controller failure in SDN, this paper proposes a novel controller offload solution for failure mitigation based on a prediction module that anticipates the presence of a harmful long-term load. In fact, the long-standing load would eventually overwhelm the controller leading to a possible failure. To predict whether the load in the controller is short-term or long-term load, we used three different classification algorithms: Support Vector Machine, k-Nearest Neighbors, and Naive Bayes. Our evaluation results demonstrate that Support Vector Machine algorithm is applicable for detecting the type of load with an accuracy of 97.93% in a real-time scenario. Besides, our scheme succeeded to offload the controller by switching between the reactive and proactive mode in response to the prediction module output.

Recent SDN researches suffer from over-accumulation of unhealthy flow-load. Instead, we leverage the SDN controller network view to encode the end-to-end path information into the packet address. Our solution EncPath significantly reduces the flow-table size and the number of control messages. Consequently, the power consumption of network switches is in orders of magnitude less than other evaluated solutions. It also provides flow management flexibility and scalability. We compare EncPath with single and multipath routing solutions and single path solution. Also, we operated them in proactive and reactive modes. We find that EncPath flow entries in core switches in a multihomed fat-tree with 144 hosts is approximately 1000 times smaller than Equal-Cost MultiPath (ECMP) and random routing. Additionally, the number of control messages to setup the network is reduced by a factor of 200×. This, consequently, affords data-plane and control-plane devices space to process other tasks.