FR3104781A1 - Device for detecting fake accounts on social networks - Google Patents

Abstract

The present invention relates to a device (1) for detecting false accounts on social networks (11), characterized in that it comprises: An extraction module (2) comprising, or configured to call on APIs (21 ) to extract data from said accounts to be verified, A prediction module comprising at least one program implementing an artificial neural network (4) constituting a machine learning model, configured to analyze said data extracted from said accounts to make predictions to provide information on the veracity of an account, or a list of false followers of an account, by pattern detection, the prediction being based on the patterns detected in the data of said fake accounts, the extracted data being data relating to the account of the user, his publications, his interactions with other members of the social network and / or his time line. Figure to be published with the abstract: Fig. 1

Description

Device for detecting fake accounts on social networks

Technical field of the invention

The present invention generally relates to the field of robot detection devices, in particular fake accounts on social networks from computer robots.

prior art

There are a large number of software "bots" that automatically perform actions of various types, for useful or harmful purposes.

The terms decrawler, crawl robot, spider, or scrapers (collecteurs in French) designate in the computer world an indexing robot, software that automatically explores the Web. It is generally designed to collect resources (web pages, images, videos, Word, PDF or PostScript documents, etc.), in order to allow a search engine to index them. Working on the same principle, some malicious bots (spambots) are used to archive resources or collect email addresses to send emails to.

A computer bot is an automatic or semi-automatic software agent that interacts with computer servers. A bot connects and interacts with the server like a client program used by a human, hence the term "bot", which is an apheresis contraction of "robot".

Social listening software (Social listening or Social media measurement) calculates the popularity of a theme or brand, and identifies emotions relating to this theme from data collected on the various media (social networks , blogs, websites, etc.). They carry out an analysis of this data in order to extract qualitative and quantitative information relating to the feeling, or the interest.

Thus, a 2018 study by Arizona State University & Carnegie Mellon University shows that 24% of all tweets are generated by bots.

Several prior art documents propose solutions to detect bots on the Internet, and several of these solutions incorporate Artificial Intelligence (AI) and Machine Learning.

WO2019063389 teaches the identification of bots/non-bots on a website by the processing of web requests, the classification of robots "bots" into several previously determined categories (associated with a degree of confidence), by the use of an AI to identify the source of web requests received. Thus, the data used in the Machine Learning model of this document is built around the web history that led to the website in question. This solution therefore requires obtaining a sufficient web history to detect the bot.

US2011131652 teaches the denial of access to a service for a "bot" robot by the use of a prediction service, based on a set of unwanted previous access data analyzed by an AI (Machine Learning) and a database of data including a history of the data. The system thus allows the real-time analysis and detection of certain indexing robots (crawlers bots, scrapers) and then their access to servers is prohibited. Thus an indexing robot (Crawler/Scraper) collects information from complete internet pages in order to store them in a database of large capacity to store all these pages. These indexing robots will come more or less regularly to the pages to update this information. These indexing robots will feed on the HTML structure (used to build a site) in order to retrieve the title by the <title> title tag or the description by the metadata <meta> tag. This solution is totally different on several levels (technical and objective) from that of the invention.

US2018152465 teaches the detection of a bot by a vector analyzed by an AI (with supervised ML models). The vector is on an idiosyncratic score related to botnet activity. In other words, the vector is based on a score linked to the actions of each virtual machine, and their individual behavior, devoid or not of potentially human actions.

WO2010143152 teaches the detection of a bot by response indicating that it is a browser (comparison of the response to a question (page request identifier (PRID)) with respect to the expected response to define whether the correspondent (Internet client) is a web crawler or not.This simple process is used to screen, or filter, the most basic bots (crawlers).

Robots present on social networks can be more complex to detect than crawlers or scrapers, and are not detectable by said prior art documents.

The object of the present invention is therefore to propose a device for detecting false accounts on social networks, making it possible to overcome at least some of the drawbacks of the prior art.

The present invention makes it possible to detect bots and networks of bots within social networks, in order to improve their readability, by improving “social listening” studies. Indeed, many of the data from social networks, and therefore analyzed by social listening platforms, come from bots, i.e. automated agent programs capable of interacting with computer servers. There is therefore a need to better identify bot networks, to be able to know if a profile is a bot or not in order to be able to calculate the true popularity of an object or a brand. These issues must be resolved with reliable and automatable solutions, which operate on a large scale (large amount of data, Big Data) and in real time.

• Un module d’extraction connecté au réseau internet et comprenant des API pour extraire des données desdits comptes à vérifier,
• Un module de prédiction comprenant au moins un ensemble de codes formant un ou plusieurs programmes, le(s) programme(s) implémentant un réseau neuronal artificiel constituant un modèle d’apprentissage automatique, de préférence par des modèles supervisés ou non supervisés, configuré pour analyser lesdites données extraites desdits comptes pour réaliser des prédictions pour fournir des informations sur la véracité d’un compte, une liste des faux comptes sur un hashtag / keyword ou une liste des faux followers, abonnés en français, d’un compte, le réseau neuronal artificiel étant entrainé sur des données de faux comptes pour détecter des motifs (patterns), la prédiction étant basée sur les motifs (patterns) détectés dans les données desdits faux comptes,

et en ce que les données extraites sont des données concernant le compte de l’utilisateur, ses publications, ses interactions avec d’autres membres du réseau social et/ou sa ligne de temps.This object is achieved by a device for detecting false accounts on social networks, the device comprising at least code modules executable on one or more processing hardware components such as a microprocessor, and a database, the device being characterized in that it comprises:

• An extraction module connected to the internet network and comprising APIs to extract data from said accounts to be verified ,
• A prediction module comprising at least one set of codes forming one or more programs, the program(s) implementing an artificial neural network constituting an automatic learning model, preferably by supervised or unsupervised models, configured to analyze said data extracted from said accounts to make predictions to provide information on the veracity of an account, a list of fake accounts on a hashtag / keyword or a list of fake followers, subscribers in French, of an account, the network artificial neural being trained on data from fake accounts to detect patterns, the prediction being based on the patterns (patterns) detected in the data from said fake accounts,

and in that the data extracted is data concerning the user's account, his publications, his interactions with other members of the social network and/or his timeline.

According to one feature, the device further comprises a module for receiving contact details of an account to be verified or of a list of accounts to be verified from one or more social networks.

According to another feature, the artificial neural network is trained, in an automated manner, on fake account data from at least one database, preferably from the database of the device, the fake account data being classified manually from accounts already identified as fake accounts.

According to another feature, the device further comprises a recording module to store the data of said accounts to be verified and false accounts in the database of the device.

According to another feature, the device further comprises a client request platform comprising an interface and a request module configured to send a request for accounts to be verified comprising at least one account identifier to be verified that the extraction module processes to extract the data of said account.

According to another feature, the module takes into account the data of the additional false accounts detected and recorded in the database of the device, to improve its predictions.

According to another feature, the device further comprises a statistics module configured to extract, sort and/or present the data from the detected accounts, for example the bot/human ratio, and/or the number of bots per day for a hashtag.

According to another feature, the device further comprises a transmitter module configured to send the results of the analysis, the list of analyzed accounts or data related to the accounts to be verified to a remote recipient.

According to another feature, the predictions of the prediction module are made at least in part according to the results from the data already present in the database.

According to another feature, the device makes predictions on randomly chosen accounts without said extraction module having a list of accounts to verify.

The invention also relates to a social listening platform, comprising social listening software capable of performing calculations on data from the social network in order to highlight trends, or to retrieve qualitative or quantitative results, the social listening further comprising a device for detecting fake accounts on social networks as described previously, the platform being capable of disregarding or focusing on data from the fake accounts present in the database of the device.

• Extraction de données d’un compte à vérifier par le module d’extraction à partir de coordonnées de compte à vérifier, d’une liste de comptes à vérifier ou de comptes choisis aléatoirement;
• Optionnellement pré-traitement des données et calcul de variables supplémentaires,
• Traitement des données par le réseau neuronal pré-entrainé du module de prédiction pour réaliser des prédictions pour fournir des informations sur la véracité d’un compte, une liste des faux comptes sur un hashtag / keyword ou une liste des faux followers, abonnés en français, d’un compte ;
• Ajout de la prédiction du compte auxdites données et enregistrement et historisation des données dans la base de données du dispositif par le module d’enregistrement.

The invention also relates to a method for detecting false accounts on social networks by a device as described above, in which the method comprises the following steps:

• Extraction of data from an account to be verified by the extraction module from account details of an account to be verified, from a list of accounts to be verified or from randomly chosen accounts;
• Optionally pre-processing of data and calculation of additional variables,
• Data processing by the pre-trained neural network of the prediction module to make predictions to provide information on the veracity of an account, a list of fake accounts on a hashtag / keyword or a list of fake followers, subscribers in French , of an account;
• Addition of count prediction to said data and recording and logging of data in the device database by the recording module.

According to another feature, the method includes an additional step of statistical analysis, sorting and/or presentation of said data analyzed by the statistical module.

• Création d’une requête par un client sur une plateforme de requête client, par l’intermédiaire d’une interface et d’un module de requête;
• Envoi de ladite requête par le module de requête au module de réception du dispositif de détection;
• Réception par le module de réception,de coordonnées de compte à vérifier ou d’une liste de comptes à vérifier, puis envoi desdites coordonnées de compte à vérifier ou de ladite liste de comptes à vérifier au module d’extraction.

According to another feature, the method comprises an additional step:

• Creation of a request by a customer on a customer request platform, through an interface and a request module;
• Sending of said request by the request module to the reception module of the detection device;
• Received by the module reception,account details to verify or a list of accounts to verify, then sending said account details to verify or said list of accounts to verify to the extraction module.

According to another feature, the method further comprises a step of training on the updated database, following the recording of the data of the new account(s) analysed.

Brief description of figures

Other characteristics, details and advantages of the invention will become apparent on reading the following description with reference to the appended figures, which illustrates:
schematically represents the device for detecting fake accounts on social networks according to certain embodiments.

detailed description of the invention

Many combinations can be envisaged without departing from the scope of the invention; the person skilled in the art will choose one or the other depending on the economic, ergonomic, dimensional or other constraints that he will have to respect.

• Un module d’extraction (2) connecté au réseau internet et comprenant des API (21) pour extraire des données desdits comptes à vérifier
• Un module de prédiction comprenant au moins un ensemble de codes formant un ou plusieurs programmes, le(s) programme(s) implémentant un réseau neuronal artificiel (4) constituant un modèle d’apprentissage automatique, de préférence par des modèles supervisés ou non supervisés, configuré pour analyser lesdites données extraites desdits comptes pour réaliser des prédictions sur la véracité du compte et indiquer pour chaque compte de la liste de faux compte reçue pour traitement si ledit compte est un faux compte, le réseau neuronal artificiel (4) étant entrainé sur des données de faux comptes pour détecter des motifs (patterns), la prédiction étant basée sur les motifs (patterns) détectés dans les données desdits faux comptes,

et en ce que les données extraites sont des données concernant le compte de l’utilisateur, ses publications, ses interactions avec d’autres membres du réseau social et/ou sa ligne de temps.In general, the present invention comprises a device (1) for detecting fake accounts on social networks (11), the device (1) comprising at least code modules executable on one or more processing hardware components such as a microprocessor, and a database (3), the device (1) further comprising:

• An extraction module (2) connected to the internet network and comprising APIs (21) for extracting data from said accounts to be verified
• A prediction module comprising at least one set of codes forming one or more programs, the program(s) implementing an artificial neural network (4) constituting an automatic learning model, preferably by supervised or unsupervised models , configured to analyze said data extracted from said accounts to make predictions on the veracity of the account and to indicate for each account of the list of false accounts received for processing whether said account is a false account, the artificial neural network (4) being trained on fake account data for detecting patterns, the prediction being based on the patterns (patterns) detected in the data of said fake accounts,

and in that the data extracted is data concerning the user's account, his publications, his interactions with other members of the social network and/or his timeline.

By “fake accounts” we mean accounts created by robot(s) automatically, and not by an individual human user. A robot, “bot” connects and interacts with the server of a social network like a client program used by a human, to automatically or semi-automatically create an account.

In some embodiments, the extracted data is data regarding the user's account, posts, interactions with other members of the social network and/or timeline, or other additional variables, such as for example account creation information.

In some embodiments, the device (1) further comprises a module for receiving contact details of an account to be verified or of a list of accounts to be verified from one or more social networks.

In some embodiments, the fake account detection implemented by the artificial neural network (4) is performed by searching for patterns in the fake account data. The detection is therefore not carried out by the creation of an idiosyncratic score, the comparison and/or the study of an answer to a question or even the processing of Web requests and the classification of bots into previously determined categories associated with a degree of confidence like what is achieved today for the detection of more basic bots such as crawlers and scrapers.

The artificial neural network comprises for example a plurality of algorithms, for example 8 Machine Learning and Deep learning algorithms. In some embodiments, some algorithms are based only on profile data and others only on the timeline. This allows a clear improvement in detection given the number of variables used. This also helps to guide the user of the device to know if the account studied is a fake account or an account powered by automated posts (publications).

The training of the neural network is thus carried out and automated from at least one database containing already identified fake accounts and their data, and preferably at least one database has been created manually specifically for this purpose. The neural network's algorithms allow it to spot patterns in the values of the various criteria and variables available for each account. This then allows him to be able to make reliable predictions on an account in order to identify it as a “False” or “Robot”, “True” or “Human”, or “Undetermined” account.

The device (1) has a software architecture that revolves around an Artificial Intelligence (AI), capable of learning automatically (by supervised or unsupervised ML models, in order to detect patterns, or patterns in French, making it possible to detect bots and networks of bots, from data of at least one account to be verified, for example resulting from the extraction by data APIs following a request sent by a user, a customer or a platform social listening (12) via a dedicated platform. Calculations can optionally be performed on said data before predictions of account statuses and logging of said data and status are made. Indeed, the device ( 1) uses APIs (21) to retrieve profile data on the various social networks (11) from the raw information of the accounts to be verified which have been provided to it, either in batch, that is to say in a list, by a platform or a client, or individually by a client via the client request platform (6) and its interface (60) comprising a request module (61). The prediction result can be displayed directly on the interface (60), or sent to the client in another way.

The extraction of data from an account is done from public data shared by said accounts, such as profile data, timeline (wall), contacts, actions, images, publications…

In some embodiments, the artificial neural network is trained, in an automated manner, on fake account data from at least one database, preferably from the database (3) of the device (1), the data from the fake accounts being manually classified from accounts already identified as fake accounts.

Thus, from accounts detected and classified manually as being fake accounts, the neural network of the device is automatically trained in order to improve its algorithms by Machine Learning mechanisms, and Deep Learning, in order to be able to subsequently make predictions. on accounts to analyze.

In some embodiments, the device (1) further comprises a module for receiving information or lists of accounts from a social network that must be verified. Accounts are generally identified by their nickname, but it is possible to identify them from a specific identifier or address.

In certain embodiments, the extraction module (2) is capable of extracting data from accounts on a website, via API (application programming interface or application programming interface, in French) . Advantageously, the extraction module makes API calls in order to retrieve the account data to be verified on one or more social networks (11).

In certain embodiments, the extraction module (2) is able to extract data sent from a social listening platform (12) from a list of accounts to be verified, preferably in batch mode.

In some embodiments, the device further comprises a registration module to store the data of said accounts to be verified and fake accounts in the database (3) of the device (1).

In some embodiments, the device (1) further comprises a client request platform (6) comprising an interface (60) and a request module (61) configured to issue a request for accounts to be verified comprising at least one identifier account to verify that the extraction module is processing to extract the data from said account.

In certain embodiments, the module takes into account the data of the additional false accounts detected and recorded in the database (3) of the device (1), to improve its predictions.

In some embodiments, the device (1) further comprises a statistics module configured to extract, sort and/or present the data from the detected accounts.

In some embodiments, the device (1) further comprises a sender module configured to send the results of the analysis, the list of analyzed accounts or data related to the accounts to be verified to a remote recipient.

In certain embodiments, the predictions of the prediction module are made at least in part according to the results from the data already present in the database (3).

In certain embodiments, the neural network of the device is configured to analyze networks of bots acting in groups, the artificial neural network (4) preferably consisting of unsupervised models.

In certain embodiments, the accounts are chosen from unique identifiers for each account, for example a nickname, from a keyword resulting from their publication, from a common member that they follow or by which they are followed, or even randomly without said extraction module (2) having a list of accounts to check.

In certain embodiments, the device (1) is part of a social listening platform (12), comprising social listening software (12) able to perform calculations on data from the social network in order to evidence of trends, or to retrieve qualitative or quantitative results, the social listening platform (12) further comprising said device (1) for detecting false accounts on social networks (11) as described previously, the platform being capable of disregarding or concentrating on the data of the false accounts present in the database (3) of the device (1).

• Extraction de données d’un compte à vérifier par le module d’extraction (2) à partir de coordonnées de compte à vérifier, d’une liste de comptes à vérifier ou de comptes choisis aléatoirement;
• Optionnellement pré-traitement des données et calcul de variables supplémentaires,
• Traitement des données par le réseau neuronal pré-entrainé du module de prédiction pour réaliser des prédictions pour fournir des informations sur la véracité d’un compte, une liste des faux comptes sur un hashtag / keyword ou une liste des faux followers, abonnés en français, d’un compte ;
• Ajout de la prédiction du compte auxdites données et enregistrement et historisation des données dans la base de données (3) du dispositif (1) par le module d’enregistrement.

The invention also relates to a method for detecting fake accounts on social networks (11) by a device (1) as described previously, in which the method comprises the following steps:

• Extraction of data from an account to be verified by the extraction module (2) from coordinates of account to be verified, from a list of accounts to be verified or from randomly selected accounts;
• Optionally pre-processing of data and calculation of additional variables,
• Data processing by the pre-trained neural network of the prediction module to make predictions to provide information on the veracity of an account, a list of fake accounts on a hashtag / keyword or a list of fake followers, subscribers in French , of an account;
• Addition of the account prediction to said data and recording and archiving of the data in the database (3) of the device (1) by the recording module.

By “pre-trained artificial neural network of the prediction module” is meant that the artificial neural network is trained on data of false accounts from a database, preferably from the database (3) of the device (1), the fake account data being manually classified as fake account data. This pre-training step allows the prediction module to make correct predictions when it is provided with real account data whose veracity we want to know.

In certain embodiments, the method comprises an additional step of statistical analysis, sorting and/or presentation of said data analyzed by the statistics module.

• Réception par le module de réception, de coordonnées de compte à vérifier ou d’une liste de comptes à vérifier, puis envoi desdites coordonnées de compte à vérifier ou de ladite liste de comptes à vérifier au module d’extraction (2).

In some embodiments, the method includes an additional step:

• Received by the module of receipt, account details to be verified or a list of accounts to be checked, then sending said account details to be checked or said list of accounts to be checked to the extraction module (2).

• Création d’une requête par un client sur une plateforme de requête client (6), par l’intermédiaire d’une interface (60) et d’un module de requête (61)
• Envoi de ladite requête par le module de requête (61) au module de réception du dispositif de détection (1)

In some embodiments, the reception step by the reception module is preceded by the following steps:

• Creation of a request by a customer on a customer request platform (6), via an interface (60) and a request module (61)
• Sending of said request by the request module (61) to the reception module of the detection device (1)

In certain embodiments, the method further comprises a step of training on the updated database (3), following the recording of the data of the new account(s) analysed.

Alternatively, the training database is completed manually every 6 months.

In certain embodiments, as for example illustrated in a nonlimiting manner in FIG. 1, a social listening platform (12) wishes to carry out social listening and monitoring of trends on social networks (11). The platform (or any third party) therefore transmits a list of accounts to be verified to the device (1) for detecting false accounts on social networks (11), which can save them in the database (3) of the device (1 ). The database (3) is preferably a decentralized cloud database (Cloud). From the list of accounts to be verified, the extraction module (2) of the device (1) connects to the social networks (11) and extracts the data corresponding to each account in the different social networks (11). This data extraction may concern profile data, the timeline (wall) of an account, or any other information public and/or shared by said account. The extraction module (2) can use application programming interfaces (21), or API (21) (application programming interface), to retrieve said data from the site or the web page of the social network. Once the data has been retrieved and stored in the database (3), the artificial neural network (41) of the prediction module (4) analyzes said extracted data to decide on the veracity of the account. It thus makes a prediction and indicates if the account is a “real” account, in other words if it is a human user who is responsible for writing the extracted content, or if the account is a “fake” account, in other words an account created and maintained by a computer robot. The prediction module can then associate the status of the account, for example true or false concerning its veracity, with the account to be verified, and/or with said extracted data, and store the result in its database (3). The social listening platform is shown here separated from the device (1), but in some embodiments, the device is integrated into the social listening platform (12).

In some embodiments, the prediction module (4) is pre-trained on data from a database of fake accounts, manually classified as such, until it has a sufficient probability of recognizing a false account during the analysis of classified false accounts, before being used during the use of the device (1) to detect false accounts among a list of accounts whose status and veracity are not known. It can further in some embodiments be retrained on the database (3) of the device (1) when new counts are detected and evaluated as false, in order to improve future detections and predictions of the artificial neural network (41) of the prediction module (4).

In some embodiments, the whole solution is hosted on the cloud (API call script, Database, device including the neural network, query platform and client interface).

In certain embodiments, the device uses, among other things, “Text mining” technology and more particularly Regex and the analysis of similarity between the last ten publications as well as the analysis of the number of languages used in the publications.

The device represents a stable model for all bots (in several fields of activity) adapting to the technological evolution of bots thanks in particular to:

The learning base which can be fed regularly (capturing technological developments) manually or automatically;

Additional algorithms can be added (we speak of standard engineering features for all users but also specific for certain users (derived algorithms)). For example, for a luxury brand, it is possible to create an additional variable “the account is present in Thailand or not”. It may therefore also be necessary to create specific variables for certain types of bots (invented in the future for example) which would not be captured by today's algorithm(s);

Unsupervised algorithms (operating as "anomaly detection", the algorithm will detect patterns (patterns) not previously identified / learned and therefore new.

In addition, the possibility of reinforcement learning (also known as reinforcement algorithm) makes it possible to automate learning during the use of the device, the neural network refining its algorithms by taking as additional data the designated counts like fake accounts.

The device also has a stable architecture, that is to say without latency, where thousands of accounts can be recovered in one go. Indeed, today, there are a lot of restrictions on APIs (on the number of posts, number of account consultations or even on the number of followers that can be analyzed over limited time windows). Scripts have therefore been designed with “sleep” functions which allow a queue to be created when the limit is reached. The advantage is not to consult the same account twice, which would happen with a “random” (random) selection of accounts requested twice by the user.

To obtain a “latency-free” device architecture, the intermediaries are preferably limited (call, store, display, analysis). It is also possible to play on the storage: store as little data as possible to make reading by the interface faster. Finally, the use of high-capacity virtual machines makes it possible to obtain very low or even non-existent latency.

One of the improvements consists in parallelizing API calls and parallelizing the creation of new variables and then the prediction. The first point involves the possession of multiple connection tokens to make calls on several different accounts at the same time.

Name Description Calculation id_str ID screen_name Nickname identifying the account name Username created_at Account creation date description Account Biography followers_count Number of followers friends_count Number of followings statuses_count Number of tweets (including retweets) favourites_count Number of tweets “liked” by the user account_age Account age in days Date today – Date created ff_ratio Ratio between the number of followings and the square of the number of followers friends_count / followers_count² reputation Ratio between the number of followers and the sum of the number of followers and the number of followings followers_count / (followers_count + friends_count) activity Average number of daily tweets since account creation statuses_count / account age impact Logorithm of the number of followers log(followers_count+1) following_rate Ratio between the number of followers and the number of followings followers_count/friends_count nb_url_description Number of urls in the description nb_hashtag_description Number of hashtags in the description alpha_ratio_screen_name Alpha character rate (letters) in nickname listed_count Number of lists attached to this account nb_name Number of terms in username protected Non-public account (approve or denied) verified Account certified by Twitter (set of criteria) geo_enabled Active geolocation has_time_zone Time zone set has_description Presence of a description (biography) has_location Presence of a location has_url The user has entered a url in the account profile has_phone_nb_description Presence of phone number in the bio is_listed Membership in a list

Name Description Calculation commitment Sum of the number of likes and retweets of the extracted nb_tweet (user's tweets only) nb_tweet Number of Tweets retrieved n_lang Number of languages used in extracted tweets mean_retweet_count Average number of retweets mean_favorite_count Average number of “liked” tweets retweet_rate Retweet rate reply_rate Response rate tweet_rate Tweet rate quote_rate Quote rate mean_url Average number of URLs mean_hashtag Average number of hashtags mean_mention Average number of mentions (@) phone_nb_rate Phone number rate among extracted tweets nb_tw_interval_inf_10s Nb tweets with a duration less than 10s rate_tw_interval_inf_10s Tweet rate with duration less than 10s tweet_hour_rate Tweet rate per hour calculated on extracted tweets similarity_rate Average similarity rate between 2 successive tweets (=> improvement over all the extracted tweets)

By way of example, and in a non-limiting way, the tables Table 1 and Table 2 respectively describe the variables User (User) and Timeline (Timeline, or antecronological wall) which can be taken into account by the neural network of the device of detection of bots on the Twitter social network. Thus, the neural network analysis module analyzes data from the account profile, its news feed, both its publications and its interactions with other social network accounts. In a particular embodiment, this data is retrieved by APIs from the identifier of an account, for example a nickname. Calculations on the data are likely to be performed to obtain more complex or more easily usable data. Credentials can be sent individually or as a batch (list).

In certain embodiments, this identifier is sent to the device from a customer request platform (6) which comprises an interface (60) on which a customer can enter said identifier and request the status of this account. The client request platform then issues a search request including the identifier of the account to be verified. Other requests are also possible, such as checking accounts that have published on a particular trend (a hashtag for example), with or without a time criterion.

The data can for example be stored in .csv in the cloud.

The request platform (6) can for example operate with Flask™ or Shiny™, preferably Shiny™.

Thus the device and the method described differ from the prior art in that the inputs and outputs (input – output) of the algorithms of the prior art differ from the algorithm or algorithms implemented in the neural network of the present invention.

• Informations de création de compte
• Informations basées sur la timeline / fil d’actualité
• Des variables supplémentaires (par des calculs réalisés sur les données récupérées).

Indeed, the invention uses the following data:

• Account creation information
• Information based on the timeline / newsfeed
• Additional variables (by calculations performed on the retrieved data).

The differences of the invention:
The training base built on existing bases and manually created bases;
The variables used / created (calculated);
Consideration of account data and newsfeed data; The aggregation of several algorithms in the neural network. In all 8 algorithms in the device;
the outputs or specific results "Output" such as: Human, Robot, indeterminate and in particular allowing the fight against misinformation.

The biggest difficulty solved by the invention is to have a stable model for all bots (in several fields of activity (luxury brands, French / European institutions)):
- personal marketing
- sale of products (vendors)
- spreading rumours, misinformation
- distribution of malware
- spamming of inappropriate content

Thus, whatever the domains, the device has a stable architecture (without latency, thousands of accounts can be recovered at once) and makes it possible to adapt to the technological evolution of bots, for example by updating the database. training on a regular basis.

It will readily be understood from reading the present application that the features of the present invention, as generally described and illustrated in the figures, can be arranged and designed in a wide variety of different configurations. Thus, the description of the present invention and the accompanying figures are not intended to limit the scope of the invention but merely represent selected embodiments.

Those skilled in the art will understand that the technical characteristics of a given embodiment can in fact be combined with characteristics of another embodiment unless the reverse is explicitly mentioned or it is obvious that these characteristics are incompatible. Moreover, the technical characteristics described in a given embodiment can be isolated from the other characteristics of this mode unless the reverse is explicitly mentioned.

It should be apparent to those skilled in the art that the present invention permits embodiments in many other specific forms without departing from the scope defined by the scope of the appended claims, they should be considered by way of illustration and the invention should not be limited to the details given above.

1. Device for detecting fake accounts
2. Extraction module
21. Data extraction API (profiles and timelines)
3. Cloud database (cloud)
4. Prediction module
41. Artificial Neural Network
5. JSON module
6. Customer Query Platform
60. Interface
61. Query Module
10.
11. Social networks
12. Social Listening Platform

Claims (16)

Device (1) for detecting fake accounts on social networks (11), the device (1) comprising at least code modules executable on one or more processing hardware components such as a microprocessor, and a database ( 3), the device (1) being characterized in that it comprises:

• An extraction module (2) connected to the Internet network and comprising, or configured to call on, APIs (21) to extract data from said accounts to be verified ,
• A module of prediction comprising at least one set of codes forming one or more programs, the program(s) implementing an artificial neural network (4) constituting an automatic learning model, preferably by supervised or unsupervised models, configured to analyze said data extracted from said accounts to make predictions to provide information on the veracity of an account, a list of fake accounts on a hashtag / keyword or a list of fake followers, subscribers in French, of an account, the network artificial neural (4) being trained on fake count data to detect patterns, the prediction being based on the detected patterns in said fake count data,

and in that the data extracted is data concerning the user's account, his publications, his interactions with other members of the social network and/or his timeline. Device (1) according to the preceding claim, in which the device (1) further comprises a module for receiving contact details of an account to be verified or of a list of accounts to be verified from one or more social networks (11) . Device (1) according to any one of the preceding claims, in which the artificial neural network is trained, in an automated manner, on data of false accounts from at least one database, preferably from the database (3) of the device (1), the data of the false accounts being classified manually from accounts already identified as being false accounts. Device (1) according to any one of the preceding claims, further comprising a registration module for storing the data of said accounts to be verified and false accounts in the database (3) of the device (1). Device (1) according to any one of the preceding claims, in which the device (1) further comprises a client request platform (6) comprising an interface (60) and a request module (61) configured to issue a request of accounts to be verified comprising at least one account identifier to be verified that the extraction module processes to extract the data from said account. Device (1) according to any one of the preceding claims, in which the extraction module (2) is able to extract data sent from a social listening platform (12) from a list of accounts to check, preferably in batch mode. Device (1) according to any one of the preceding claims, in which the module takes into account the data of the additional fake accounts detected and recorded in the database (3) of the device (1), to improve its predictions. Device (1) according to any one of the preceding claims, in which the device (1) further comprises a statistics module configured to extract, sort and/or present the data from the detected accounts. Device (1) according to any one of the preceding claims, in which the device (1) further comprises a transmitter module configured to send the results of the analysis, the list of accounts analyzed or data related to the accounts to be verified to a remote recipient. Device (1) according to any one of the preceding claims, in which the predictions of the prediction module are made at least in part according to the results from the data already present in the database (3). Device (1) according to any one of the preceding claims, in which a random generator of account identifiers to be verified allows predictions to be made on randomly chosen accounts without said extraction module (2) having a list accounts to verify. Social listening platform (12), comprising social listening software (12) capable of performing calculations on data from the social network in order to highlight trends, or to retrieve qualitative or quantitative results, the social listening (12) being characterized in that it further comprises or communicates with a JavaScript Object Notation (JSON) JSON (5) module translating the internet format into a textual data format derived from the JavaScript language object notation adapted to the device (1) for detecting false accounts on social networks (11) according to any one of the preceding claims, the platform being capable of disregarding or of concentrating on data from the false accounts present in the database (3) of the device (1). Method for detecting fake accounts on social networks (11) by a device (1) according to any one of claims 1 to 11, in which the method comprises the following steps:

• Extraction of data from an account to be verified by the extraction module (2) from coordinates of account to be verified, from a list of accounts to be verified or from randomly chosen accounts;
• Pre-processing of data and optional calculation of additional variables,
• Data processing by the pre-trained neural network of the prediction module to make predictions to provide information on the veracity of an account, a list of fake accounts on a hashtag / keyword or a list of fake followers, subscribers in French , of an account;
• Addition of the prediction of the account to said data and recording and archiving of the data in the database (3) of the device (1) by the recording module.

Method for detecting false accounts according to the preceding claim, in which the method comprises an additional step of statistical analysis, sorting and/or presentation of said data analyzed by the statistics module. A method of detecting fake accounts according to any one of claims 13 or 14, wherein the method comprises an additional step:

• Creation of a request by a customer on a customer request platform, via an interface (60) and a request module (61)
• Sending of said request by the request module (61) to the reception module of the detection device (1)
• Received by the module of receipt, account details to be verified or a list of accounts to be checked, then sending said account details to be checked or said list of accounts to be checked to the extraction module (2).

Method for detecting false accounts according to any one of Claims 13 to 15, in which the method also comprises a step of training on the updated database (3), following the recording of the data of the new analyzed accounts.