IT202000021253A1 - IT platform based on artificial intelligence systems to support IT security - Google Patents

Description

Description of the industrial invention entitled:

?IT platform based on artificial intelligence systems to support IT security?

The invention that we are going to present takes advantage of artificial intelligence solutions, unique in the world, which when combined in an appropriate manner with others already developed. existing, within an IT platform based on artificial intelligence systems - which will later be? identified as a Security Box - allow you to solve problems of security and privacy of information.

Problem

For some time now, and more and more? in the future, companies of all sectors and sizes purchase services from external companies, services that increasingly? are based on IT solutions which in turn are ?hosted? on cloud services of other companies, often abroad. The approach of decentralizing services ? more and more inevitable, but brings with it? a number of security issues.

Although no system (not just information technology) can? never be totally immune to attacks, especially large multinationals have achieved high levels of protection of their data centers and personal data.

To date, however, there are no good systems to protect the chain of (outsourced) services.

Not ? in fact, it is sufficient to protect the information and one's own services with the highest levels of sophistication that the technology allows if then people outside the company, to whom? been entrusted with the management of them, they can easily access and operate (Fig.1).

Assuming that the service itself? is protected with the highest security standards, nothing prevents an external actor from altering it, voluntarily or involuntarily (human error of internal personnel), using vulnerabilities? not necessarily of an IT nature (e.g. social engineering).

The idea

Sistem-Evo srl, with a research project in the field of artificial intelligence, has developed an innovative and unique reasoning mechanism, inspired by the functioning of the human brain, which can be replicated by any industry expert.

This advanced reasoning mechanism, inserted inside the Security Box, allows you to define very complex rules of behavior and management of information. The Security Box, owned by of? user and in its exclusive management, will allow? to also take advantage of external third-party services, maintaining absolute control over alterations, anomalies, attacks that may occur (Fig. 2).

With the flow and the designed architecture, each internal and external service must be ?filtered? from the Security Box in which the ?reasonings? have been loaded, thanks to the innovative invention called HBMS (Human Brain Modeling System) which will allow to validate the passage of information.

HBM extension

The HBMS reasoning system, loaded inside the Security Box, is an artificial intelligence inspired by the logic of combining concepts, which lead to reasoning, and therefore deduction, typical of the human brain (Fig.3).

L? HBMS exploits a large graph structure where each node represents an elementary concept. Any concept? connected with another through one or more? weights through a mechanism that implements logical operators, which allow the? activation of a concept through the propagation of one or more? signals that exceed a threshold.

The HBMS ? unprecedented of its kind in that its structure simultaneously serves as a knowledge base and an engine of inference.

Why? a security box

The user, who has exclusive management of the Security Box, can? model both very detailed and high-level generic reasoning / deductions.

Rules can also be drawn using natural language interpretation logic (where needed or expected, but not mandatory)

By way of example:

? If today? Saturday or Sunday do not accept changes to the service from external sources (third parties).

All services, internal and external, are inputs of the Security Box, which by validating them according to the reasoning rules it contains, will block them? or less.

Given the flexibility? of the HBMS reasoning system these inputs can be heterogeneous in nature.

By way of example:

? Video

? Images

? text documents

? Text field content of a transaction from API ? Virtual Assistants

Technical Insights

The component of A.I. at the base of this solution it exploits innovative solutions that compensate for the intrinsic limits of machine learning / deep learning.

The mathematical structure underlying all ML/DL models (Fig. 4 - Topology of a neural network) always provides for a level called input, zero or more? intermediate levels, and an output level, all formed by a variable number of nodes (model topology).

Net of some variations, such as the dropout technique, each level is ? hyper-connected, i.e. each node ? connected with all nodes of the previous level.

This structure allows the various variations of Machine Learning and Deep Learning algorithms to solve problems and carry out highly complex tasks, in some cases to exceed the capabilities of the human being, however they have structural limitations that make them absolutely ineffective in other scenarios, one of these ? the one described in this document in relation to the Security Box.

Let's see exactly what are the limits we are talking about and how the proposed solution manages to get around them.

Unpredictability of the model

There are mathematical theorems that demonstrate how it is impossible to predict whether a mathematical model is so? did it converge? or less, or in summary if the mathematical model we designed will work? or less and even more? in how much time of training this will be able? to happen.

This limit means that we have no guarantees that we will always be able to set ?rules? functioning with extreme certainty.

Rigid structure

The topology of an ML/DL model must be determined in advance; any modification, however minimal, involves retraining the mathematical model again, meeting the problems identified in the section ?Unpredictability? of the model?.

In the illustrated scenario of the Security Box, this feature makes it impossible, even at the most? sophisticated ML/DL algorithm, update the security rules it contains in real time.

Expensive training

Due to the nature of the ML/DL models, the training process provides for a number more? or lower number of examples from which ?to learn? the task they will have to perform. Pi? in detail their aim is to find the coefficients of a system of equations which approximate the objective function.

Furthermore, to work at its best, the totality? of the examples given in training must be balanced on the number of expected output classes, i.e. an equal number of examples for each category expected to classify.

A similar scenario? often unfeasible for a variety of reasons, but very often because? the necessary data simply do not exist or are they ?unbalanced? in number, e.g. 1000 examples for category A (Those who will buy), 100 examples for category B (Those who won't buy) causing the problem that the model will learn? to classify only in case A.

As the complexity increases of the model also increases exponentially, the need? computational in terms of hardware; this also involves an increase in costs, not always sustainable, taking into account what is highlighted in the sections ?Unpredictability? of the model?, ?Rigid structure?, ?Expensive training?. In a scenario like the one described for the Security Box, it is not Is it possible to foresee each new safety rule that will be able to? be inserted, carry out training and count on the fact that it will go? successful.

Advantages and uniqueness of AI HBMS

Sistem-Evo has invested many resources in research and development to overcome the limitations described above, while maintaining the benefits and typical power of ML/DL.

The result ? was the invention of a new neural network with a graph structure in which the nodes represent the neurons and the arcs (relationships) represent the connections between them.

The graph structure has an important difference with respect to the topology described for the M/DL models:

? There are no levels of input, intermediate or output;

? Neurons are not hyperconnected, connections between nodes exist only where expected.

Each neuron in the neural network has a ?purpose? and a specific semantic meaning and can be combined in an extremely targeted way (remember that in ML/DL models, all neurons are connected indiscriminately to those of the previous level). The combination of neurons generates a neuron which, for convenience, we will call a ?concept?. The combination of more neurons and concepts generates a sub-graph which culminates in a node called, for convenience, ?inference?, effectively forming an ?reasoning?.

This neural network design implies that those identified as input nodes are selected in real time based on the ?query? which is carried out.

In fact it is extracted the pi? small sub-graph which, starting from the inputs activated by the query, leads to a ?reasoning? in common with them.

This kind of new approach eliminates all the problems identified in the ?Unpredictability? of the model?, ?Rigid structure?, ?Expensive training?.

The model, taking advantage of the property? of the graphs, ? in fact extremely dynamic as we can modify and update the neural network in real time without having to retrain the entire model. Major appearance? that the neural network designed by Sistem-Evo not only ? predictable (contrary to the ML/DL models, with what follows, see section ?Rigid Structure?) but ? possible to control the behavior of each individual neuron, which is absolutely unique in the field of A.I.

This graph architecture, together with the adoption of complex numbers, has allowed the introduction of further innovation, the possibility of to give a semantic meaning to the neurons, or the ?concepts?.

Thanks to the possibility to be able to control the combination of pi? neurons (that is, modify the weights of the connections between neurons in order to determine whether or not they can be activated to propagate the signal to the subsequent ones), ? possible to construct a ?reasoning? (or a combination of neurons according to a graph structure) much more? similar to the mechanisms of the human brain than to the behavior of a mathematical model (Fig. 5).

Fields of application

The fields in which a neural network like the one proposed can excel are mainly NLP (Natural Language Processing) or interpretation of natural language and the definition of reasoning rules for deductive purposes.

On these fields, the neural network in question has brilliantly outperformed all the mathematical models of ML/DL currently known.

Concrete example of application

Sistem-Evo has developed various products in the A.I. including a virtual assistant, named IVO, which takes full advantage of the newly developed neural network technology described herein.

In addition to excelling in capabilities? NLP, the particular graph structure of the network has allowed the development of unique solutions of their kind or shared and multilevel knowledge bases.

Specifically, Sistem-Evo has developed an idea that is based on a shared knowledge base, which contains all the Italian (or foreign) language, applicable to all customers. There are, then, further basic levels of knowledge that are more? specialized in the field of use, such as, for example, the specific knowledge base for Hotels, Restaurants etc. which overlaps with the previous generic knowledge base.

In the end, ? expected a level that contains and defines the specific knowledge base of the customer, which ? valid only for it.

Sistem-Evo and its partners hold control of all levels of knowledge bases as well as? of the answers associated with each inference (for simplicity, the answers associated with the possible questions that can be asked to the virtual assistant).

In these circumstances, the user does not therefore have control over the contents that are managed in his name and will be able to do so. be counted only after these have possibly been altered by editing errors of the delivery group.

more scenarios however, complex and delicate transactions may concern customers operating in the banking sector. In this case, the user wants to have guarantees on the contents that are published in his name, managed by third parties.

Sistem-Evo has therefore further strengthened its service by offering the Security Box solution which integrates perfectly into this scenario; the answers to possible questions are not immediately provided to the end user by the Sistem-Evo computer systems, but are first sent to the security box.

The customer, who himself manages the security box within his structure - to which Sistem-Evo has no access - can? define rules, such as the following, in natural language: ?prevent posting of racially discriminatory content? if, purely by way of example, the external team, but also internal where foreseen, associated a ?question? content such as: ?Our company offers specific cosmetic products for the hair of ?black? people.

If this way of expressing yourself does not conform to the wishes, to the communication rules modeled inside the Security Box, this message would be blocked by the latter which would undertake one or more actions. foreseen actions, such as, for example, sending an alert.

This kind of problem is often underestimated by non-experts, but a source of great concern for those who work with the public.

To give a better idea of the extent of the problem that the Security Box solves, let's introduce what in the literature is called Machine Learning Bias.

In recent years, Machine Learning has brought innovation and undisputed benefits, but because of its nature? was also the cause of serious documented and therefore verifiable facts.

Artificial intelligence, specifically ML/DL, ? been used in advanced solutions to support personnel selection.

Since, as known, these systems ?learn? from the historical data in possession, in different scenarios - one of all the technological ones, but also happened in the government sector where women have a lower employment ratio - the A.I. has in fact learned to discriminate against women in the pre-analysis phase in the recruiting process; likewise, it has learned to discriminate against men in jobs that are mostly represented by women.

The Bias in the ML / DL are very difficult to manage and, in almost all? some cases, the problems that derive from it, such as those previously described, are never wanted, but often we realize this by chance, after a long time, by which time we have in the face of the irreparable.

The Security Box in this scenario, even supporting the existing ML/DL systems (remember that it is not invasive towards the currently installed systems) could have prevented all this from happening.

Being a ?White Box?, the user could have instructed ?reasoning? to intercept and regulate anomalous behaviors such as those provided in the example.

The computer literature is already? for some time highlighting the dangers of Black Box algorithms such as the known ML / DL ones and the consequent Bias.

Conclusions

To date, all efforts related to computer security are concentrated on curbing intrusion attempts or attacks aimed at corporate or private computer systems. As far as data protection is concerned, however, can we? rely on information encryption algorithms. The Security Box solution is aimed at unprecedented computer security scenarios, for which nothing is? still been done or the control of? integrity? of information, protection against voluntary or involuntary tampering with data and company knowledge entrusted to partners or third-party companies.

The Security Box managed centrally (within its corporate structure) and educated through a set of rules and high-level reasoning that only the user can define and control, carries out this work in an extremely efficient way, relying on an artificial intelligence which, in turn, is unique of its kind, managing to overcome all the limits that the use of Machine Learning / Deep Learning algorithms would impose, instead, on the functionalities? of the Security Box.

Claims (3)

CLAIMS Claims of the industrial invention entitled: ?IT platform based on artificial intelligence systems to support IT security? 1. IT platform for the creation of a device characterized by the fact that it includes a hardware infrastructure within which loaded and executed: - an artificial intelligence software module, characterized by the fact that it includes a neural network whose components are based on systems of rules and reasoning typical of the human brain aimed at intercepting security anomalies relating to the integrity of the exchange of information between systems and to undertake actions as programmed, providing partial or total control to the operator who manages it without the need? of new trainings in the face of data changes (updating, deleting, inserting), - a graph database that represents the relationships between the elements that form the security rules defined by the user, thus representing the general knowledge base, - a semantic text analysis software module aimed at extracting all the grammatical elements of a sentence in addition to the main topic (here defined as semantic subject), - a speech to text software module, aimed at receiving an audio file as input and converting it into a textual representation, - a text to speech software module, aimed at receiving a text format as input and converting it into a playable audio format, - a graphical interface and command line through which the user can? interact with the database for maintenance which includes: ? a software module used for the definition, modification, cancellation of the individual elements that make up the rules, according to one or more? of the preceding claims, ? a software module used to manage the relationships between the single elements of the rules, i.e. the modification of the weights (complex or real numbers in floating point), ? a software module aimed at generating new rules autonomously from the analysis of a textual or vocal input through the use of the semantic analysis module according to one or more of the preceding claims, - a software module dedicated to the collection of statistical usage data, processed by the artificial intelligence module which, also on the basis of rules set by the user, according to one or more of the preceding claims, allows to improve the performance of the system by relating the data. 2. IT platform, according to one or more? of the preceding claims, characterized in that it comprises hardware sensors capable of sending audio, video, environmental data inputs (for example, temperature, pressure) which, according to one or more previous claims, feed the inputs of the artificial intelligence module through: - A software module which, using computer vision algorithms supported by machine learning algorithms, powered by one or more? hardware modules operate biometric analysis for facial identification and vital parameters aimed at security. 3. IT platform, according to one or more? of the preceding claims, characterized in that it comprises a software module which, through machine learning algorithms, uses wearable devices connected to hardware sensors which analyze biometric parameters aimed at identifying the person and, therefore, at safety, i.e. they feed the module of artificial intelligence referred to in claim 1).