Posts

cryptographie, nombres aléatoires, random numbers

Cryptography: what are the random numbers for?

/0 Comments/in , /by

Hervé Debar, Télécom SudParis – Institut Mines-Télécom and Olivier Levillain, Télécom SudParis – Institut Mines-Télécom

The original purpose of cryptography is to allow two parties (traditionally referred to as Alice and Bob) to exchange messages without another party (traditionally known as Eve) being able to read them. Alice and Bob will therefore agree on a method to exchange each message, M, in an encrypted form, C. Eve can observe the medium through which the encrypted message (or ciphertext) C is sent, but she cannot retrieve the information exchanged without knowing the necessary secret information, called the key.

This is a very old exercise, since we speak, for example, of the ‘Julius Caesar Cipher’. However, it has become very important in recent years, due to the increasing need to exchange information. Cryptography has therefore become an essential part of our everyday lives. Besides the exchange of messages, cryptographic mechanisms are used in many everyday objects to identify and authenticate users and their transactions. We find these mechanisms in phones, for example, to encrypt and authenticate communication between the telephone and radio antennas, or in car keys, and bank cards.

The internet has also popularized the ‘padlock’ in browsers to indicate that the communication between the browser and the server are protected by cryptographic mechanisms. To function correctly, these mechanisms require the use of random numbers, the quality (or more precisely, the unpredictability) thereof contributes to the security of the protocols.

Cryptographic algorithms

To transform a message M into an encrypted message C, by means of an algorithm A, keys are used. In so-called symmetric algorithms, we speak of secret keys (Ks), which are shared and kept secret by Alice and Bob. In symmetric algorithms, there are public (KPu) and private (KPr) key pairs. For each user, KPu is known to all, whereas KPr must be kept safe by its owner. Algorithm A is also public, which means that the secrecy of communication relies solely on the secrecy of the keys (secret or private).

Sometimes, the message M being transmitted is not important in itself, and the purpose of encrypting said message M is only to verify that the correspondent can decrypt it. This proof of possession of Ks or KPr can be used in some authentication schemes. In this case, it is important never to use the same message M more than once, since this would allow Eve to find out information pertaining to the keys. Therefore, it is necessary to generate a random message NA, which will change each time that Alice and Bob want to communicate.

The best known and probably most widely used example of this mechanism is the Diffie-Helman algorithm.  This algorithm allows a browser (Alice) and a website (Bob) to obtain an identical secret key K, different for each connection, by having exchanged their respective KPu beforehand. This process is performed, for example, when connecting to a retail website. This allows the browser and the website to exchange encrypted messages with a key that is destroyed at the end of each session. This means that there is no need to keep it (allowing for ease of use and security, since there is less chance of losing the key). It also means that not much traffic will be encrypted with the same key, which makes cryptanalysis attacks more difficult than if the same key were always used.

Generating random numbers

To ensure Eve is unable obtain the secret key, it is very important that she cannot guess the message NA. In practice, this message is often a large random number used in the calculations required by the chosen algorithm.

Initially, generating random variables was used for a lot of simulation work. To obtain relevant results, it is important not to repeat the simulation with the same parameters, but to repeat the simulation with different parameters hundreds or even thousands of times. The aim is to generate numbers that respect certain statistical properties, and that do not allow the sequence of numbers to be differentiated from a sequence that would be obtained by rolling dice, for example.

To generate a random number NA that can be used in these simulations, so-called pseudo-random generators are normally used, which apply a reprocessing algorithm to an initial value, known as the ‘seed’.  These pseudo-random generators aim to produce a sequence of numbers that resembles a random sequence, according to these statistical criteria. However, using the same seed twice will result in obtaining the same sequence twice.

The pseudo-random generator algorithm is usually public. If an attacker is able to guess the seed, he will be able to generate the random sequence and thus obtain the random numbers used by the cryptographic algorithms. In the specific case of cryptography, the attacker does not necessarily even need to know the exact value of the seed. If they are able to guess a set of values, this is enough to quickly calculate all possible keys and to crack the encryption.

In the 2000s, programmers used seeds that could be easily guessed, that were based on time, for example, making systems vulnerable. Since then, to avoid being able to guess the seed (or a set of values for the seed), operating systems rely on a mixture of the physical elements of the system (e.g. processing temperature, bus connections, etc.). These physical elements are impossible for an attacker to observe, and vary frequently, and therefore provide a good seed source for pseudo-random generators.

What about vulnerabilities?

Although the field is now well understood, random number generators are still sometimes subject to vulnerabilities. For example, between 2017 and 2021, cybersecurity researchers found 53 such vulnerabilities (CWE-338). This represents only a small number of software flaws (less than 1 in 1000). Several of these flaws, however, are of a high or critical level, meaning they can be used quite easily by attackers and are widespread.

A prime example in 2010 was Sony’s error on the PS3 software signature system. In this case, the reuse of a random variable for two different signatures allowed an attacker to find the manufacturer’s private key: it then became possible to install any software on the console, including pirated software and malware.

Between 2017 and 2021, flaws have also affected physical components, such as Intel Xeon processors, Broadcom chips used for communications and Qualcom SnapDragon processors embedded in mobile phones. These flaws affect the quality of random number generation.  For example, CVE-2018-5871 and CVE-2018-11290 relate to a seed generator whose periodicity is too short, i.e. that repeats the same sequence of seeds quickly. These flaws have been fixed and only affect certain functions of the hardware, which limits the risk.

The quality of random number generation is therefore a security issue. Operating systems running on newer processors (less than 10 years old) have random number generation mechanisms that are hardware-based. This generally ensures a good quality of the latter and thus the proper functioning of cryptographic algorithms, even if occasional vulnerabilities may arise. On the other hand, the difficulty is especially prominent in the case of connected objects, whose hardware capacities do not allow the implementation of random generators as powerful as those available on computers and smartphones, and which often prove to be more vulnerable.

Hervé Debar, Director of Research and Doctoral Training, Deputy Director, Télécom SudParis – Institut Mines-Télécom and Olivier Levillain, Assistant Professor, Télécom SudParis – Institut Mines-Télécom

This article has been republished from The Conversation under a Creative Commons license. Read the original article.

MuTAS, urban mobility

“En route” to more equitable urban mobility, thanks to artificial intelligence

/0 Comments/in , , /by

Individual cars represent a major source of pollution. But how can you transition from using your own car when you live far from the city center, in an area with little access to public transport? Andrea Araldo, researcher at Télécom SudParis is undertaking a research project that aims to redesign city accessibility, to benefit those excluded from urban mobility.

The transport sector is responsible for 30% of greenhouse gas emissions in France. And when we look more closely, the main culprit appears clearly: individual cars, responsible for over half of the CO2 discharged into the atmosphere by all modes of transport.

To protect the environment, car drivers are therefore thoroughly encouraged to avoid using their car, instead opting for a means of transport that pollutes less. However, this shift is impeded by the uneven distribution of public transport in urban areas. Because while city centers are generally well connected, accessibility proves to be worse on the whole in the suburbs (where walking and waiting times are much longer). This means that personal cars appear to be the only viable option in these areas.

The MuTAS (Multimodal Transit for Accessibility and Sustainability) project, selected by the National Research Agency (ANR) as part of the 2021 general call for projects, aims to reduce these accessibility inequalities at the scale of large cities. The idea is to provide the keys to offering a comprehensive, equitable and multimodal range of mobility options, combining public transport with fixed routes and schedules with on-demand transport services, such as chauffeured cars or rideshares. These modes of transport could pick up where buses and trains leave off in less-connected zones. “In this way, it is a matter of improving accessibility of the suburbs, which would allow residents to leave their personal car in the garage and take public transport, thereby contributing to reducing pollution and congestion on the roads”, says Andrea Araldo, researcher at Télécom SudParis and head of the MuTAS project, but formerly a driving school owner and instructor!

Improving accessibility without sending costs sky-high

But how can on-demand mobility be combined with the range of public transport, without leading to overblown costs for local authorities? The budget issue remains a central challenge for MuTAS. The idea is not to deploy thousands of vehicles on-demand to improve accessibility, but rather to make public transport more equitable within urban areas, for an equivalent cost (or with a limited increase).

This means that many questions must be answered, while respecting this constraint. In which zones should on-demand mobility services be added? How many vehicles need to be deployed? How can these services be adapted to different times throughout the day? And there are also questions regarding public transport. How can bus and train lines be optimized, to efficiently coordinate with on-demand mobility? Which are the best routes to take? Which stations can be eliminated, definitively or only at certain times?

To resolve this complex optimization issue, Araldo and his teams have put forward a strategy using artificial intelligence, in three phases.

Optimizing a graph…

The first involves modeling the problem in the form of a graph. In this graph, the points correspond to bus stops or train stations, with each line represented by a series of arcs, each with a certain trip time. “What must be noted here is that we are only using real-life, public data,” emphasizes Araldo. “Other research has been undertaken around these issues, but at a more abstract level. As part of MuTAS, we are using openly available, standardized data, provided by several cities around the world, including routes, schedules, trip times etc., but also population density statistics. This means we are modeling real public transport systems.” On-demand mobility is also added to the graph in the form of arcs, connecting less accessible areas to points in the network. This translates the idea of allowing residents far from the city center to get to a bus or train station using chauffeured cars or rideshares.

L’attribut alt de cette image est vide, son nom de fichier est Schema-1024x462.png.
To optimize travel in a certain area, researchers start by modeling public transport lines with a graph.

…using artificial intelligence

This modeled graph acts as the starting point for the second phase. In this phase, a reinforcement learning algorithm is introduced, a method from the field of machine learning. After several iterations, this is what will determine what improvements need to be made to the network, for example, deactivating stations, eliminating lines, adding on-demand mobility services, etc. “Moreover, the system must be capable of adapting its structure dynamically, according to shifts in demand throughout the day,” adds the researcher. “The traditional transport network needs to be dense and extended during peak hours, but it can contract significantly in off-peak hours, with on-demand mobility taking over for the last kilometers, which is more efficient for lower numbers of passengers.”

And that is not the only complex part. Various decisions influence each other: for example, if a bus line is removed from a certain place, more rideshares or chauffeured car services will be needed to replace it. So, the algorithm applies to both public transport and on-demand mobility. The objective will therefore be to reach an optimal situation in terms of equitable distribution of accessibility.

But how can this accessibility be evaluated? There are multiple methods to do so, but researchers have chosen two adapted methods for graph optimization. The first is a ‘velocity score’, corresponding to the maximum distance that can be traveled from a departure point in a limited time (30 minutes for example). The second is a ‘sociality score’, representing the number of people that one can meet from a specific area, also within a limited time.

In concrete terms, the algorithm will take an indicator as a reference, i.e. a measure of the accessibility for the least accessible place in the area. The aim being to make transport options as equitable as possible, it will aim to optimize this indicator (‘max-min’ optimization), while respecting certain restrictions such as cost. To achieve this, it will make a series of decisions concerning the network, initially in a random way. Then, at the end of each iteration, by analyzing the flow of passengers, it will calculate the associated ‘reward’, the improvement in the reference indicator. The algorithm will then stop when the optimum is reached, or else after a pre-determined period.

This approach will allow it to establish knowledge of its environment, associating each network structure (according to the decisions made) with the expected reward. “The advantage of such an approach is that once the algorithm is trained, the knowledge base can be used for another network,” explains Araldo. “For example, I can use the optimization performed for Paris as a starting point for a similar project in Berlin. This represents a precious time-saver compared to traditional methods used to structure transport networks, in which you have to start each new project from zero.”

Testing results on (virtual) users in Ile-de-France

Lastly, the final phase aims to validate the results obtained using a detailed model. While the models from the first phase aim to reproduce reality, they only represent a simplified version. This is important, given that they will then be used for various iterations, as part of the reinforcement learning process. If they had a very high level of detail, the algorithm would require a huge amount of computing power, or too much processing time.

The third phase therefore involves first delicately modeling the transport network in an urban area (in this case, the Ile-de-France region), still using real-life data, but more detailed this time. To integrate all this information, researchers use a simulator called SimMobility, developed at MIT in a project to which Araldo contributed. The tool makes it possible to simulate the behavior of populations at an individual level, each person represented by an ‘agent’ with their own characteristics and preferences (activities planned during the days, trips to take, desire to reduce walking time or minimize number of changes, etc.). ‎It was based on the work of Daniel McFadden (Nobel Prize for Economics in 2000) and Moshe Ben-Akiva on ‘discrete choice models’, which makes it possible to predict choices between multiple modes of transport.

With the help of this simulator and public databases (socio-demographic studies, road networks, numbers of passengers, etc.), Araldo and his team, in collaboration with MIT, will generate a synthetic population, representing Ile-de-France users, with a calibration phase. Once the model faithfully reproduces reality, it will be possible to submit it to the new optimized transport system and simulate user reactions. “It is important to always remember that it’s only a simulation,” reminds the researcher. “While our approach allows us to realistically predict user behavior, it certainly does not correspond 100% to reality. To get closer, more detailed analysis and deeper collaborations with transport management bodies will be needed.”

Nevertheless, results obtained could serve to support more equitable urban mobility and in time, reduce its environmental footprint. Especially since the rise of electric vehicles and automation could increase the environmental benefits. However, according to Araldo, “electric, self-driving cars do not represent a miracle solution to save the planet. They will only prove to be a truly eco-friendly option as part of a multimodal public transport network.”

Bastien Contreras

digital simulation

Digital simulation: applications, from medicine to energy

/0 Comments/in , /by

At Mines Saint-Étienne, Yann Gavet uses image simulation to study the characteristics of an object. This method is more economical in terms of time and cost, and eliminates the need for experimental measurements. This field, at the intersection of mathematics, computer science and algorithms, is used for a variety of applications ranging from the medical sector to the study of materials.

What do a human cornea and a fuel cell electrode have in common? Yann Gavet, a researcher in applied mathematics at Mines Saint-Étienne1 is able to model these two objects as 2D or 3D images in order to study their characteristics. To do this, he uses a method based on random fields. “This approach consists in generating a synthetic image representing a surface or a random volume, i.e. whose properties will vary from one point to another across the plane or space,” explains the researcher. In the case of a cornea, for example, this means visualizing an assembly of cells whose density differs according to whether we look at the center or the edge. The researcher’s objective? To create simulations with properties as close as possible to the reality.

Synthetic models and detecting corneal disorders

The density of cells that make up our cornea –the transparent part at the front of the eye– and its endothelium, provides information about its health. To perform these analyses, automatic cell detection and counting algorithms have been developed using deep neural networks. Training them thus requires access to large databases of corneas. The problem is that these do not exist in sufficient quantity. “However, we have shown that it is possible to perform the training process using synthetic images, i.e. simulated by models,” says Yann Gavet.

How does it work? Using deep learning, the researcher creates graphical simulations based on key criteria: size, shape, cell density or the number of neighboring cells. He is able to simulate cell arrangements, as well as complete and realistic images of corneas. However, he wants to combine the two. Indeed, this step is essential for the creation of image databases that will allow us to train the algorithms. He focuses in particular on the realism of the simulation results in terms of cell geometry, gray levels and the “natural” variability of the observations.

Although he demonstrated that training using synthetic corneal data does not require perfectly realistic representations to perform well, improving accuracy will be useful for other applications. “As a matter of fact, we transpose this method to the simulation of material arrangements that compose fuel cell electrodes, which requires more precision,” explains the researcher.

Simulating the impact of microstructures on the performance of a fuel cell

The microstructure of fuel cell electrodes impacts the performance and durability of solid oxide cells. In order to improve these parameters, researchers want to identify the ideal arrangement of the materials that make up the electrodes, i.e., how they should be distributed and organized. To do this, they play with the “basic” geometry of an electrode: its porosity and its material particle size distribution. This therefore targets the morphological parameters on which the manufacturers intervene when designing the electrodes.

To identify the best performing structures, one method would be to build and test a multitude of configurations. This is an expensive and time-consuming practice. The other approach is based on the simulation and optimization of a large number of configurations. Subsequently, a second group of models simulating the physics of a battery can in turn identify which structures best impact the battery’s performance.

The advantage of the simulations is that they target specific areas within the electrodes to better understand their operation and their overall impact on the battery. For example: exchange zones such as “triple phase” points where ionic, electronic and gaseous phases meet, or exchanges between material surfaces. “Our model allows us to evaluate the best configuration, but also to identify the associated manufacturing process that offers the best energy efficiency for the battery,” says Yann Gavet.

In the medium term, the researcher wishes to continue his work on a model whose dimensions are similar to the observations made in X-ray tomography. An algorithmic challenge that will require more computing time, but will also lead to results that are closer to the reality of the field.

1 Yann Gavet is a researcher at the Georges Friedel laboratory, UMR CNRS/Mines Saint-Étienne

Anaïs Culot

Comprendre informations du langage, algorithms

Making algorithms understand what we are talking about

/0 Comments/in , /by

Human language contains different types of information. We understand it all unconsciously, but explaining it systematically is much more difficult. The same is true for machines. The NoRDF Project Chair “Modeling and Extracting Complex Information from Natural Language Text” seeks to solve this problem: how can we teach algorithms to model and extract complex information from language? Fabian Suchaneck and Chloé Clavel, both researchers at Telecom Paris, explain the approaches of this new project

What aspects of language are involved in making machines understand?

Fabian Suchaneck: We need to make them understand more complicated natural language texts. Current systems can understand simple statements. For example, the sentence: “A vaccine against Covid-19 has been developed” is simple enough to be understood by algorithms. On the other hand, they cannot understand sentences that go beyond a single statement, such as: “If the vaccine is distributed, the Covid-19 epidemic will end in 2021. In this case, the machine does not understand that the condition required for the Covid-19 epidemic to end in 2021 is that the vaccine is distributed. We also need to make machines understand what emotions and feelings are associated with language; this is Chloé Clavel’s specialist area.

What are the preferred approaches in making algorithms understand natural language?

FS: We are developing “neurosymbolic” approaches, which seek to combine symbolic approaches with deep learning approaches. Symbolic approaches use human-implemented logical rules that simulate human reasoning. For the type of data we process, it is fundamental to be able to interpret what has been understood by the machine afterwards. Deep learning is a type of automatic learning where the machine is able to learn by itself. This allows for greater flexibility in handling variable data and the ability to integrate more layers of reasoning.

Where does the data you analyze come from?

FS: We can collect data when humans interact with chatbots from a company and especially those from the project’s partner companies. We can extract data from comments on web pages, forums and social networks.

Chloé Clavel: We can also extract information about feelings, emotions, social attitudes, especially in dialogues between humans or humans with machines.

Read on I’MTech: Robots teaching assistants

What are the main difficulties for the machine in learning to process language?

CC: We have to create models that are robust in changing contexts and situations. For example, there may be language variability in the expression of feelings from one individual to another, meaning that the same feelings may be expressed in very different words depending on the person. There is also a variability of contexts to be taken into account. For example, when humans interact with a virtual agent, they will not behave in the same way as with a human, so it is difficult to compare data from these different sources of interactions. Yet, if we want to move towards more fluid and natural human-agent interactions, we must draw inspiration from the interactions between humans.

How do you know whether the machine is correctly analyzing the emotions associated with a statement?

CC: The majority of the methods we use are supervised. The data entered into the models are annotated in the most objective way possible by humans. The goal is to ask several annotators to annotate the emotion they perceive in a text, as the perception of an emotion can be very subjective. The model is then taught about the data for which a consensus among the annotators could be found. When testing the performance of the model, when we inject an annotated text into a model that has been trained with similar texts, we can see if the annotation it produces is close to those determined by humans.

Since the annotation of emotions is particularly subjective, it is important to determine how the model actually understood the emotions and feelings present in the text. There are many biases in the representativeness of the data that can interfere with the model and mislead us on the interpretation made by the machine. For example, if we assume that younger people are angrier than older people in our data and that these two categories do not express themselves in the same way, then it is possible that the model may end up simply detecting the age of the individuals and not the anger associated with the comments.

Is it possible that the algorithms end up adapting their speech according to perceived emotions?

CC: Research is being conducted on this aspect. Chatbots’ algorithms must be relevant in solving the problems they are asked to solve, but they must also be able to provide a socially relevant response (e.g. to the user’s frustration or dissatisfaction). These developments will improve a range of applications, from customer relations to educational or support robots.

What contemporary social issues are associated with the understanding of human language by machines?

FS: This would notably allow a better understanding of the perception of news on social media by humans, the functioning of fake news, and therefore in general which social group is sensitive to which type of discourse and why. The underlying reasons why different individuals adhere to different types of discourse are still poorly understood today. In addition to the emotional aspect, there are different ways of thinking that are built in argumentative bubbles that do not communicate with each other.

In order to be able to automate the understanding of human language and exploit the numerous data associated with it, it is therefore important to take as many dimensions into account as possible, such as the purely logical aspect of what is said in sentences and the analysis of the emotions and feelings that accompany them.

By Antonin Counillon

Eclairer boites noires, algorithms

Shedding some light on black box algorithms

/0 Comments/in , /by

In recent decades, algorithms have become increasingly complex, particularly through the introduction of deep learning architectures. This has gone hand in hand with increasing difficulty in explaining their internal functioning, which has become an important issue, both legally and socially. Winston Maxwell, legal researcher, and Florence d’Alché-Buc, researcher in machine learning, who both work for Télécom Paris, describe the current challenges involved in the explainability of algorithms.

What skills are required to tackle the problem of algorithm explainability?

Winston Maxwell: In order to know how to explain algorithms, we must draw on different disciplines. Our multi-disciplinary team, AI Operational Ethics, focuses not only on mathematical, statistical and computational aspects, but also on sociological, economic and legal aspects. For example, we are working on an explainability system for image recognition algorithms used, among other things, for facial recognition in airports. Our work therefore encompasses these different disciplines.

Why are algorithms often difficult to understand?

Florence d’Alché-Buc: Initially, artificial intelligence used mainly symbolic approaches, i.e., it simulated the logic of human reasoning. Logical rules, called expert systems, allowed artificial intelligence to make a decision by exploiting observed facts. This symbolic framework made AI more easily explainable. Since the early 1990s, AI has increasingly relied on statistical learning, such as decision trees or neural networks, as these structures allow for better performance, learning flexibility and robustness.

This type of learning is based on statistical regularities and it is the machine that establishes the rules which allow their exploitation. The human provides input functions and an expected output, and the rest is determined by the machine. A neural network is a composition of functions. Even if we can understand the functions that compose it, their accumulation quickly becomes complex. So a black box is then created, in which it is difficult to know what the machine is calculating.

How can artificial intelligence be made more explainable?

FAB: Current research focuses on two main approaches. There is explainability by design where, for any new constitution of an algorithm, explanatory output functions are implemented which make it possible to progressively describe the steps carried out by the neural network. However, this is costly and impacts the performance of the algorithm, which is why it is not yet very widespread. In general, and this is the other approach, when an existing algorithm needs to be explained, it is an a posteriori approach that is taken, i.e., after an AI has established its calculation functions, we will try to dissect the different stages of its reasoning. For this there are several methods, which generally seek to break the entire complex model down into a set of local models that are less complicated to deal with individually.

Why do algorithms need to be explained?

WM: There are two main reasons why the law stipulates that there is a need for the explainability of algorithms. Firstly, individuals have the right to understand and to challenge an algorithmic decision. Secondly, it must be guaranteed that a supervisory institution such as the  French Data Protection Authority (CNIL), or a court, can understand the operation of the algorithm, both as a whole and in a particular case, for example to make sure that there is no racial discrimination. There is therefore an individual aspect and an institutional aspect.

Does the format of the explanations need to be adapted to each case?

WM: The formats depend on the entity to which it needs to be explained: for example, some formats will be adapted to regulators such as the CNIL, others to experts and yet others to citizens. In 2015, an experimental service to deploy algorithms that detect possible terrorist activities in case of serious threats was introduced. For this to be properly regulated, an external control of the results must be easy to carry out, and therefore the algorithm must be sufficiently transparent and explainable.

Are there any particular difficulties in providing appropriate explanations?

WM: There are several things to bear in mind. For example, information fatigue: when the same explanation is provided systematically, humans will tend to ignore it. It is therefore important to use varying formats when presenting information. Studies have also shown that humans tend to follow a decision given by an algorithm without questioning it. This can be explained in particular by the fact that humans will consider from the outset that the algorithm is statistically wrong less often than themselves. This is what we call automation bias. This is why we want to provide explanations that allow the human agent to understand and take into consideration the context and the limits of algorithms. It is a real challenge to use algorithms to make humans more informed in their decisions, and not the other way around. Algorithms should be a decision aid, not a substitute for human beings.

What are the obstacles associated with the explainability of AI?

FAB: One aspect to be considered when we want to explain an algorithm is cyber security. We must be wary of the potential exploitation of explanations by hackers. There is therefore a triple balance to be found in the development of algorithms: performance, explainability and security.

Is this also an issue of industrial property protection?

WM: Yes, there is also the aspect of protecting business secrets: some developers may be reluctant to discuss their algorithms for fear of being copied. Another counterpart to this is the manipulation of scores: if individuals understand how a ranking algorithm, such as Google’s, works, then it would be possible for them to manipulate their position in the ranking. Manipulation is an important issue not only for search engines, but also for fraud or cyber-attack detection algorithms.

How do you think AI should evolve?

FAB: There are many issues associated with AI. In the coming decades, we will have to move away from the single objective of algorithm performance to multiple additional objectives such as explainability, but also equitability and reliability. All of these objectives will redefine machine learning. Algorithms have spread rapidly and have enormous effects on the evolution of society, but they are very rarely accompanied by instructions for their use. A set of adapted explanations must go hand in hand with their implementation in order to be able to control their place in society.

By Antonin Counillon

Also read on I’MTech: Restricting algorithms to limit their powers of discrimination

 

David Gesbert, PERFUME

PERFUME: a scent of cooperation for the networks of the future

/0 Comments/in , /by

The ERC PERFUME project, led by EURECOM researcher David Gesbert and ending in 2020, resulted in the development of algorithms for local decision making in the mobile network. This research was tested on autonomous drones, and is particularly relevant to the need for connected robotics in the post-5G world.

Now that 5G is here, who’s thinking about what comes next? The team working with David Gesbert, a researcher specializing in wireless communication systems at EURECOM, has just completed its ERC PERFUME project on this subject. So what will wireless networks look like by 2030? While 5G is based on the centralization of calculations in the cloud, the networks of the future will require, on the contrary, a distributed network. By this, we mean the emergence of a more cooperative network. “In the future, the widespread use of robotic objects and devices to perform autonomous tasks will increase the need for local decision making, which is difficult in a centralized system,” says Gesbert. Nevertheless, the objective remains the same: optimizing the quality of the network. This is especially important since the increase in connected devices may cause more interference and therefore affect the quality of the information exchanged.

Why decentralize decision making on the network?

Under 5G, every device that is connected to the network can send measurements to the cloud. The cloud has a very high computing capacity, enabling it to process an immeasurable amount of data, before sending instructions back to devices (a tablet, cell phone, drone, etc.). However, these information transfers take time, which is a very valuable commodity for connected robotics applications or critical missions. Autonomous vehicles, for example, must make instant decisions in critical situations. “In the context of real-time applications, the response speed of the network must be optimized. Decentralizing decisions closer to the base stations is precisely the solution that was studied in our PERFUME project,” explains David Gesbert. As 5G is not yet equipped to meet this constraint, we have to introduce new evolutions of the standard.

EURECOM’s researchers are thus relying on cooperation and coordination of the computing capabilities of local terminals such as our cell phones. By exchanging information, these terminals could coordinate in the choice of their power and transmission frequency, which would limit the interference that would limit the flow rates, for example. They would no longer focus solely on their local operations, but would participate in the overall improvement of the quality of the network. A team effort that would manifest itself at the user level by sending files faster or providing better image quality during a video call. However, although possible, this collaboration remains difficult to implement.

Towards more cooperative wireless networks

Distributed networks pose a major problem: access to information from one device to another is incomplete. “Our problem of exchanging information locally can be compared to a soccer team playing blindfolded. Each player only has access to a noisy piece of information and doesn’t know where the other team members are in their attempt to score the goal together”, says David Gesbert. Researchers then develop so-called robust decision-making algorithms. Their objective? To allow a set of connected devices to process this noisy information locally. “Networks have become too complicated to be optimized by conventional mathematical solutions, and they are teeming with data. This is why we have designed algorithms based on signal processing but also on machine learning,” continues the researcher.

These tools were then tested in a concrete 5G network context in partnership with Ericsson. “The objective was for 5G cells to coordinate on the choice of directional beams of MIMO (multi-input multi-output) antennas to reduce interference between them,” says the researcher. These smart antennas, deployed as part of 5G, are increasingly being installed on connected devices. They perform “beamforming”, which means that they direct a radio signal in a specific direction – rather than in all directions – thus improving the efficiency of the signal. These promising results have opened the door to large-scale tests on connected robotics applications, the other major focus of the ERC project. EURECOM has thus experimented with certain algorithms on autonomous drones.

Drones at the service of the network?

Following a disaster such as an avalanche, a tsunami or an earthquake, part of the ground network infrastructure may be destroyed and access to the network may be cut off. It would then be possible to replicate a temporary network architecture on site using a fleet of drones serving as air relays. On the EURECOM campus, David Gesbert’s team has developed prototypes of autonomous drones connected to 5G. These determine a strategic flight position and their respective positions in order to solve network access problems for users on the ground. The drones then move freely and recalculate their optimal placement according to the user’s position.  This research notably received the prize for the best 2019 research project, awarded by the Provence-Alpes-Côte d’Azur region’s Secure Communicating Solutions cluster.

This solution could be considered in the context of rescue missions and geolocalization of missing persons. However, several challenges need to be addressed for this method to develop. Indeed, current regulations prohibit the theft of autonomous aircraft. In addition, they have a flight time of about 30 minutes, which is still too short to offer sustainable solutions.

“This research is also adapted to issues relating to autonomous cars, adds David Gesbert. For example, when two vehicles arrive at an intersection, a protocol for coordination must be put in place to ensure that the vehicles cross the intersection as quickly as possible and with the lowest likelihood of collision.” In addition, medicine and connected factories would also be targets for application of distributed networks. As for the integration of this type of standard in the future 6G, it will depend on the interests of industrial players in the years to come.

By Anaïs Culot

Learn more about the ERC PERFUME project