RESUMO
It is 10 years since neural networks made their spectacular comeback. Prompted by this anniversary, we take a holistic perspective on artificial intelligence (AI). Supervised learning for cognitive tasks is effectively solved-provided we have enough high-quality labelled data. However, deep neural network models are not easily interpretable, and thus the debate between blackbox and whitebox modelling has come to the fore. The rise of attention networks, self-supervised learning, generative modelling and graph neural networks has widened the application space of AI. Deep learning has also propelled the return of reinforcement learning as a core building block of autonomous decision-making systems. The possible harms made possible by new AI technologies have raised socio-technical issues such as transparency, fairness and accountability. The dominance of AI by Big Tech who control talent, computing resources, and most importantly, data may lead to an extreme AI divide. Despite the recent dramatic and unexpected success in AI-driven conversational agents, progress in much-heralded flagship projects like self-driving vehicles remains elusive. Care must be taken to moderate the rhetoric surrounding the field and align engineering progress with scientific principles.
RESUMO
Trojan backdoor is a poisoning attack against neural network (NN) classifiers in which adversaries try to exploit the (highly desirable) model reuse property to implant Trojans into model parameters for backdoor breaches through a poisoned training process. To misclassify an input to a target class, the attacker activates the backdoor by augmenting the input with a predefined trigger that is only known to her/him. Most of the proposed defenses against Trojan attacks assume a white-box setup, in which the defender either has access to the inner state of NN or is able to run backpropagation through it. In this work, we propose a more practical black-box defense, dubbed TrojDef. In a black-box setup, the defender can only run forward-pass of the NN. TrojDef is motivated by the Trojan poisoned training, in which the model is trained on both benign and Trojan inputs. TrojDef tries to identify and filter out Trojan inputs (i.e., inputs augmented with the Trojan trigger) by monitoring the changes in the prediction confidence when the input is repeatedly perturbed by random noise. We derive a function based on the prediction outputs which is called the prediction confidence bound to decide whether the input example is Trojan or not. The intuition is that Trojan inputs are more stable as the misclassification only depends on the trigger, while benign inputs will suffer when augmented with noise due to the perturbation of the classification features. Through mathematical analysis, we show that if the attacker is perfect in injecting the backdoor, the Trojan infected model will be trained to learn the appropriate prediction confidence bound, which is used to distinguish Trojan and benign inputs under arbitrary perturbations. However, because the attacker might not be perfect in injecting the backdoor, we introduce a nonlinear transform to the prediction confidence bound to improve the detection accuracy in practical settings. Extensive empirical evaluations show that TrojDef significantly outperforms the-state-of-the-art defenses and is highly stable under different settings, even when the classifier architecture, the training process, or the hyperparameters change.
RESUMO
Underwater acoustic wireless sensor networks (UAWSNs) have many applications across various civilian and military domains. However, they suffer from the limited available bandwidth of acoustic signals and harsh underwater conditions. In this work, we present an Orthogonal Frequency Division Multiple Access (OFDMA)-based Media Access Control (MAC) protocol that is configurable to suit the operating requirements of the underwater sensor network. The protocol has three modes of operation, namely random, equal opportunity and energy-conscious modes of operation. Our MAC design approach exploits the multi-path characteristics of a fading acoustic channel to convert it into parallel independent acoustic sub-channels that undergo flat fading. Communication between node pairs within the network is done using subsets of these sub-channels, depending on the configurations of the active mode of operation. Thus, the available limited bandwidth gets fully utilized while completely avoiding interference. We derive the mathematical model for optimal power loading and subcarrier selection, which is used as basis for all modes of operation of the protocol. We also conduct many simulation experiments to evaluate and compare our protocol with other Code Division Multiple Access (CDMA)-based MAC protocols.
RESUMO
The aim of this study was to evaluate the short-term and long-term results of the subadventitial resection of carotid chemodectomas and to validate the importance of deliberate resection of the external carotid artery (ECA). From 1981 to 2006, 39 carotid chemodectomas of the carotid bifurcation or of the neighboring nerves were operated on in our department. There were 14 men and 22 women, with a mean age 44.4 +/- 5 (range 21-78) years. One of these operations was a redo surgery for local recurrence. One female patient presented with a bilateral tumor. Two tumors were secreting catecholamines. All these tumors affected the carotid body; 10 of them were also affecting the vagus nerve, and one among these last 10 affected the sympathetic nerve as well. In 11 cases, the tumor had spread into the subparotidal space and, in one case, into the skull. In two cases, the tumor had been revealed by hemispheric ischemia and in every case by tumoral syndrome. All the patients were followed up by clinical examination, duplex scan, or computed tomographic scan until the end of 2006. In 38 cases, complete resection was performed; an incomplete resection was performed in one case with cranial invasion. Under general anesthesia, and most of the time without pharmaceutical preparation, surgery consisted of a deliberate sacrifice of the ECA followed by subadventitial resection of the tumor. In one case, a previous embolization had been carried out to facilitate the cleavage, which in fact rendered it more complicated. In 22 cases, resection concerned the ECA; in seven cases, it concerned the common carotid artery and the internal carotid artery (ICA): in seven cases the superior laryngeal nerve, in nine cases the vagus nerve, in five cases the sympathetic nerve, and in four cases the jugular vein. In 13 cases, node clearing was associated. In 20 cases, an additional vascular procedure was performed: nine dilatations for spasm of the ICA, five autogenous vein grafts, two prosthetic bypasses, and one endarterectomy associated with a patch angioplasty. All patients were followed up until 2006. At 3 months, the observed complications were the sequelae of a homolateral hemispheric accident due to thrombosis of a vein graft, eight peripheral facial nerve palsies, 12 vocal palsies, seven Claude Bernard-Horner (CBH) syndromes, eight palatal paralyses, and 10 nociceptive pains. Some of these complications did persist: nine vocal cord paralyses that were successfully treated by speech therapy, three mild CBH syndromes, and nociceptive pains in 6% of the cases (15.4%), incapacitating in one case. With a follow-up of 115 +/- 27 (range 1-298) months, three local recurrences were recorded at 6 and 10 years. In two cases, local recurrence occurred when initial resection of the ECA had not been performed. Two patients presented with a contralateral lesion, at 12 and 16 years, respectively. At 40 months, one patient had to be reoperated on for an atheromatous stenosis. At 51 months, a female patient's death was not related to the operation. Subadventitial resection of carotid body tumors with deliberate resection of the ECA is a simple and efficient procedure. It is the preferential treatment for these slow-growing localized tumors.
