爱立信：2023年可信AI白皮书-对电信业的影响（英文版）（20页）.pdf

1、Ericsson White PaperGFTL-23:000757 UenJune 2023Trustworthy AI-What it means for telecomTrustworthy AI-What it means for telecomContentJune 20232ContentIntroduction 3Human Agency and Oversight 4Transparency 5Privacy and Data Governance 8Diversity,Non-discrimination and Fairness 9Technical Robustness

2、and Safety 11Social and Environmental Well Being 13Conclusion 15Glossary 16References 18Authors 19Trustworthy AI-What it means for telecomIntroductionJune 20233IntroductionBillions of people have come to trust and depend on modern telecom systems to support their needs and quality of life.As these s

3、ystems adopt new technologies,its important trust is maintained by understanding and addressing any new risks.Artificial intelligence(AI)differs from traditional software in its construction and operation and may introduce new and varied risks,calling for new countermeasures and guardrails.For examp

4、le,the large amount of data used for AI training raises the possibility of privacy risks.Development procedures must ensure that AI models learn what is intended.The model operation should be thoroughly understood,for example,using explainability techniques.In short,to maintain the trustworthiness o

5、f the overall system,AI must itself be trustworthy:meaning,it should operate as intended and do no harm physically or ethically.Governments,companies,and standards bodies around the world are taking notice of these facts and creating requirements regarding the trustworthiness of AI systems.The upcom

6、ing European Union AI Act 1 is one such effort.It follows principles written by the European Commission High-Level Expert Group in their“Ethics Guidelines for Trustworthy AI”2.Ericsson has adopted these guidelines.The framework breaks trustworthiness into seven specific areas.This paper explores how

7、 six of these areas apply to AI in telecom systems,some of which are depicted in Figure 1.Figure 1:Ericssons activities and technologiesTransparency Explainable RL Explainable ML/MR Explainable GNN XAI Quantification Causal AIPrivacy&Data Governance ML for Security Security for ML Privacy Preserving

8、 AI Federated LearningTechnical Robustness&Safety Safe RL Automated model quality assurance Invariance&directional expectation tests NFL/RL for safety Formal verificationSocietal Environmental wellbeing Energy management use case AI for GoodTrustworthy AI-What it means for telecomHuman Agency and Ov

9、ersightJune 20234Human Agency and OversightHuman agency and oversight requirements make sure that humans can always intervene in AI-controlled systems before things like fundamental rights and safety have the potential of being affectedin other words before they could pose any harm.Implementation of

10、 such requirements means having“humans in the loop”,the difficulty of which depends on the timescale of the decision and the criticality of the system.Sometimes AI operates at timescales,which are much too fast for human intervention,such as the optimization of radio operations in a base station.In

11、telecom,these uses typically do not have a direct impact on the rights and safety of individuals.However,they can dramatically impact network operations,which can subsequently affect humans,so they need careful assessment.Implementing human agency and oversight in AI-aided network operations require

12、s user research,product function design,and user testing to make sure that network operation engineers can detect and intervene when needed.The human-machine interaction design must be aligned with existing network operation processes,rely on existing interfaces,and provide actionable information to

13、 users.The interface for human agency and oversight may vary depending on the use case and users.It can be a GUI,CLI,Rest-based API,or even a physical interface(for example,a light).For instance,an AI system that detects and predicts network-wide congestion may use a graphical interface embedded in

14、a dashboard used for regular network design and operations.The system may also send alerts if AI performance deteriorates(for example,by reporting too many false network congestions)to the network operations center(NOC).The alerts should include everything the NOC engineer needs,including reasons fo

15、r the problem,potential root causes,and solution suggestions.The engineer then has three choices:to switch to a non-AI based function;to understand and solve the AI issue;or to escalate it.Alert volume should be considered,to not overwhelm a potentially already heavily loaded NOC.Explainable AI meth

16、ods can help in generating needed,and user-tailored,reasons for problem.The use case,user knowledge,and persistence of the AI notification influence the actions to be taken.Intermittent or ephemeral events may need to be repeated.More serious events may need an escalation path.Trustworthy AI-What it

17、 means for telecomTransparencyJune 20235TransparencyTrust can arise from understanding how a system works,or from experience using it over time.The complexity and black-box nature of AI can lead to suspicion,particularly when people feel that the AIs own creators dont fully understand how it makes d

18、ecisions,and what exactly it has learned.Greater transparency can help build trust by understanding and explaining AI models to humans.Explainable AI(XAI)refers to methods and techniques that produce models which show why and how an AI algorithm has made a certain decision.It helps stakeholders unde

19、rstand how decisions are being made in different formats:by identifying what input factors were most important in making an inference and by providing explanations and responding to“why”and/or“what-if”questions.It also helps a human operator in decision-making.If the operator is not satisfied with a

20、 response,a further investigation can be performed,using computational argumentation techniques.Creators of AI for telecom should provide XAI methods to help build the trust of their direct customers(for example,service providers),and in turn,enable them to build trust for their subscribers.The expl

21、ainability of AI should start with design and continue through implementation,as a built-in feature,to ensure transparency throughout the AI development lifecycle.In addition,different XAI techniques should be researched,and developed to explain different types of machine learning(ML)methods.The Eri

22、csson white paper 3 presents different XAI techniques applied to different AI/ML methods,including machine reasoning(MR),and reinforcement learning(RL).The explanations generated by these XAI techniques not only help explain the decisions to humans but also support automation,for example,in root cau

23、se analysis when combined with other AI techniques.Explainability of ML,that is,feature analysis techniques(including SHAP and LIME)can be used in multiple telecom use cases to identify and explain the problems and root causes of specific ML model outputs in addition to ensuring the overall correctn

24、ess of the ML models.These techniques can be applied to ML-based predictions to investigate the most important features that contribute to certain prediction results and validate the correctness of the ML model.The results of these explainability techniques can support MR components in identifying t

25、he root cause of the problem.5G slice assurance is one such use case where these techniques are thoroughly investigated and tested.In this use case,certain Quality-of-Service(QoS)requirements(such as throughput,latency,and availability)are agreed upon with the customer in a service level agreement(S

26、LA)and must be met throughout Trustworthy AI-What it means for telecomTransparencyJune 20236the lifecycle of the slice.ML models are used to proactively identify any potential violation of the agreed QoS requirements.Upon a violation in prediction,explainability techniques are applied to identify th

27、e most contributing features which in turn helps NOC Engineers in identifying the root cause of the problem 4.These techniques can be applied to multiple use cases in a similar manner,like cell shaping and key performance indicator(KPI)degradation prediction,focusing on latency-and network throughpu

28、trelated optimizations.Explainability of RL:RL is suitable to solve many cellular network problems due to its dynamic nature,online training,interaction with the environment,and outstanding performance over traditional rule-based techniques for the telecom domain.An RL agent performs an action(such

29、as applying a policy)in an environment to maximize rewards.The explainability of RL includes methods applied to different RL components,such as rewards and policy explanations.In a base station,the antennas are tilted up,down,or kept the same to optimize KPIs,that is,the coverage of the network,incr

30、eased quality by reducing interference,and capacity/throughput of the network.Coverage refers to the area from which a UE can access the cellular network,while capacity refers to the amount of traffic the cellular network can handle simultaneously.Remote electrical tilt(RET)refers to adjusting the t

31、ilt of the antenna by an RL agent to optimize the above mentioned KPIs.Increasing the down-tilt reduces the area covered by the antenna,with the risk of leaving a certain area without coverage but increases the capacity in the covered area due to a stronger signal.In contrast,up-tilt results in a la

32、rger area covered but lower capacity due to a weaker signal.Explainability is important in the RET optimization of antennas in a cellular network.Explanations help in understanding the reasons behind a specific adjustment.The following explainable reinforcement learning(XRL)methods are applied to th

33、is use case 5:Figure 2:Total Reward is decomposed into multiple sub-reward functions for better explainabilityTotal RewardMultiple sub-rewardsTrustworthy AI-What it means for telecomTransparencyJune 20237 Reward decomposition provides intuitive contrastive local explanations for the agents decisions

34、 by decomposing the reward into multiple sub-functions to adjust the tilt(see Figures 2 and 3),while achieving the same performance as the original DQN algorithm.The generated contrastive explanations are very user-interpretable,as they concisely answer questions in the form of“why did you decide to

35、 down-tilt instead of up-tilting?”The Linear Model U-Tree(LMUT)reaches high performance while employing a fully transparent linear model capable of generating both local and global explanations(see Figure 4),however,it is less transparent than reward decomposition.Autonomous Policy Explanation summa

36、rizes the trained policy and explains it in natural language,thus enabling the policy to be understood by everyone including non-experts.Contrastive explanation through embedded self-prediction produces a local explanation about the internal representation of the RL agents(intermediate or inner laye

37、rs of the deep neural network).It compares two different actions,such as why the antenna is tilted down and not up.Figure 4:LMUT distils a transparent model from a trained agent by transferring the NN to a decision tree with linear regression in the leaf node,where the tree is inherently transparent

38、Q=22.78-3.77*RSRP-8.47*SINR+6.43*TiltQ=25.79-2.08*RSRP 2.63*SINR-3.77*TiltQ=22.3-6.0*RSRP-0.38*SINR-9.17*TiltRSRP-0.72RSRP 0.34RSRPSINRTputTiltRSRPQ-valuesIn addition to enabling transparency and AI automation,Ericsson has seen the potential of reducing the input feature set by using XRL.A novel met

39、hod was developed for connecting explanations from both the input(feature analysis)and output(reward)ends of a black-box RL model,resulting in fine-grained explanations 6.Reward prioritization,performed by the user,generates two different levels of explanation,and allows RL agent reconfigurations wh

40、en unwanted behaviors are observed.Figure 3:Reward decomposed into coverage(RSRP),quality(SINR),and capacity(throughput)sub-functions for RET use caseQ-valuesStateStateThroughputRSRPSINRQ-valuesDQNdrDQNTrustworthy AI-What it means for telecomPrivacy and Data GovernanceJune 20238Privacy and Data Gove

41、rnanceIt is usually necessary or desirable to prevent AI systems data from being disclosed.If the data includes the personal data of individuals,it may be subject to stringent legal requirements.Business data might contain intellectual property or be subject to contractual constraints.Sometimes,the

42、data or other information related to it can be inferred from an AI model,especially when combined with publicly available data sources.Laws and regulations requiring privacy for individuals,such as the EU General Data Protection Regulation(GDPR)13,predate the mass adoption of AI across industries.De

43、spite being written in a manner to make them future-proof,such regulations dont necessarily anticipate the extent of potential AI risks.There are also often contractual requirements for privacy.And even in the absence of such requirements,it is generally understood that maintaining privacy is ethica

44、lly the right thing to do.Ensuring privacy in telecom AI impacts the entire AI lifecycle and requires the application of Privacy by Design and Privacy by Default(as defined by GDPR 7 and other global privacy laws),starting with controls where the training data is collected and continuing through mod

45、el use(inference).Many controls relevant to AI are the same as or similar to those used in other types of data processing.The nature of the data,the purpose of the collection,who will use it,and how and when it will be used,should be clearly communicated.Only the minimum amount of data required for

46、the intended purpose should be collected.Unnecessary fields should be redacted or masked.Controls like pseudonymization,encryption,authentication,and authorization should be used to ensure appropriate access.Uses of the data should be logged and auditable.Once no longer needed,data should be securel

47、y erased.Since AI models learn from the training data,in some sense the data is encompassed within the model.Therefore,its unsurprising that attacks exist on models to extract the training data itself or make inferences about it.Privacy enhancing technologies(PETs),including differential privacy,exi

48、st to help make AI models less susceptible to such attacks.Trustworthy AI-What it means for telecomPrivacy and Data GovernanceJune 20239Developers can also analyze model sensitivity to data extraction,for example,how many queries are required.Such metrics can inform decisions about the privacy risk

49、involved in deploying the model.Data can be exposed during training if it is improperly secured.When federated learning is used,multiple participants with independent datasets can contribute to building a single global model.This helps them keep their datasets private,but the protocols used must be

50、carefully designed to ensure no information is leaked.Trustworthy AI-What it means for telecomDiversity,Non-discrimination and FairnessJune 202310Diversity,Non-discrimination and FairnessBias in models can come from bias in the training data,which itself may stem from historical prejudices and inequ

51、ities.It may also be caused by disproportional representation.One example is natural language processing systems used for interaction with subscribers,such as chatbots,or support ticketing systems.Even within a single language,training such systems should consider different speaking styles,idioms,an

52、d education levels.Not everyone speaks perfectly,but everyone deserves the same level of service.Steps to avoid bias include:understanding what categories exist in the input data that need to be treated equally making sure each category is adequately represented in the training set,regardless of siz

53、e being aware of historical inequities that might be relevant to the problem at hand and adjusting the data accordingly ensuring model robustness,including thorough testing considering categories separately throughout the model development process,so that the model performs well for eachEven a seemi

54、ngly solely technical problem can have a bias that impacts people,sometimes in subtle ways.Consider the use of AI to structure and operate mobile networks.This can be impacted(or biased)based on how and where that data is collected,or by cognitive bias already present in the systems creators.If data

55、 collection is skewed toward a particular group(for example,people with a certain economic status,which might correlate to other factors such as race),the resulting system might be inadvertently biased.Telecom systems have to work everywhere in the world and should provide equally good service to al

56、l people.Trustworthy AI-What it means for telecomDiversity,Non-discrimination and FairnessJune 202311If more investment(that is,equipment,optimization,effort)goes into certain areas,some of these areas might get disproportionately better service.Such areas could have different usage patterns.When ML

57、 that is used to plan,deploy,optimize,and operate networks is trained on data collected in some regions,networks(or products)might be created that work well only in those regions.Trustworthy AI-What it means for telecomTechnical Robustness and SafetyJune 202312Technical Robustness and SafetyHuman sa

58、fety is not typically directly impacted by telecom operations but can be affected by situations like loss of service.Telecom systems can be important components in emergency communications and disaster handling.Another example is an application like autonomous vehicles,where the loss of communicatio

59、n might impact the ability to proceed safely.AI can assist in these situations when it makes the network itself more robust.But this means that the AI itself must be robust.Careful attention to AI quality must be taken during training and deployment.Fallback mechanisms should be in place for cases w

60、here the AI cannot decide or makes an out-of-bounds decision to,for example,transfer the control to a human operator.AI is also subject to new types of attacks.Data extraction attacks,which have already been mentioned,compromise the training data and potential privacy.Researchers have also demonstra

61、ted attacks on inference using adversarial examples.Carefully crafted inputs are fed to the model,causing its inference to be biased in a direction chosen by the attacker.Where the attacker has access,poisoning the training data can be used to influence model operation.A motivated attacker might use

62、 these mechanisms to affect network operations or steal a service.The training pipeline,resulting models,and surrounding application context should be analyzed for susceptibility to such attacks.Automated model quality assurance is crucial for technical robustness and safety.Models must be thoroughl

63、y tested against performance metrics that reflect potential real-world scenarios.These metrics are use case and model specific.For instance,in the case of a classification model,accuracy,false positive or false negative rates might be measured,Trustworthy AI-What it means for telecomTechnical Robust

64、ness and SafetyJune 202313while in the case of clustering,Silhouette Coefficient 8 and Dunn Index 9 might be chosen.It is also important to communicate this information in an understandable and useful manner to model recipients,keeping in mind that they may not have expertise in AI.For example,when

65、communicating a performance metric about a model,information should be included about its meaning and what values are considered good or bad.Since a models performance depends on the data set used for training and evaluation,data quality is essential for model quality.A training data set should accu

66、rately represent reality and cover the events or objects of interest:the training data set must have the same statistical properties the real objects have,and if there are relationships between the attributes of the real object or among the real objects,those relationships must be preserved in the t

67、raining data set.The RET use case mentioned earlier shows that poor AI decisions could lead to interference and compromised network performance.Anywhere an AI algorithm is used in the operation of a network,a failure of that algorithm can lead to inefficiency,instability,or,at worst,downtime.Another

68、 important concern is that AI models could,inadvertently or maliciously,take actions that are unsafe for humans.As already mentioned,telecom systems by nature are not safety-critical,and the danger could be reduced network performance.One scenario could be when the model explores the space of all po

69、ssible states and actions.This becomes significantly important for RL,where space exploration is seen as an effective way to train an RL agent to capture a near-optimal policy.However,unchecked exploration can lead the system to visit a dangerous state,for example,when the system tries to tilt the a

70、ntenna at an overly high angle.Safe RL methods provide a shield to block unsafe actions that might result from free exploration of state-action spaces.The intention is to allow the agent some state exploration of the environment while having boundaries using safety specifications defined by a human

71、developer.The specifications,or boundaries,can be user dependent.One such use case where these techniques are successfully tested is RET Optimization 10.Conducting invariance and directional expectation tests is also essential to assess and assure the models robustness.In an invariance test,label-pr

72、eserving perturbations are applied to inputs and the model prediction is expected to remain the same.In a directional expectation test,a set of perturbations are made to the input which should have a predictable effect on model output.Trustworthy AI-What it means for telecomSocial and Environmental

73、well beingJune 202314Social and Environmental well beingAI can be used to create positive benefits for society,such as by helping protect the environment.Ericsson considers communication to be a fundamental human right,so the availability of the network is core to societal well-being.The findings of

74、 this white paper support this conclusion.AI,by helping service providers create telecom networks that are more reliable,ubiquitous,and inexpensive,contributes to the social goal of universal communication.But there are possible negatives,such as the privacy aspects discussed above.ML training can b

75、e energy intensive,so careful cost-benefit analysis is needed before it is employed.ML might be used where there are potential safety or societal risks(for example,control of critical infrastructure).Minimizing those risks is important,and this can be done using the techniques described in this whit

76、e paper.AI can help improve network operations and energy consumption.Large volumes of data can be used to optimize important goals like performance,reliability,capacity,and energy usage.Traditional optimizations,written by programmers,typically used only a few parameters and simple algorithms with

77、limited results.AI allows large numbers of parameters to be used,better optimizations,and therefore better performance.Ericsson teams looking at specific use cases have shown that AI can lead to significant energy savings compared to traditional algorithms.11.AI in telecom can be useful beyond the s

78、imply better operation of the network.During the COVID-19 pandemic,Ericsson engaged in a joint project with a service provider,government officials,and two hospitals 12.The service provider provided anonymized and aggregated data about peoples movements,taken from their network.This was combined wit

79、h vaccination,antibody test,and hospital COVID patient admission data.A series of Trustworthy AI-What it means for telecomSocial and Environmental well beingJune 202315ML models used the data to predict admissions two to three weeks into the future.In eleven of sixteen weeks,the predictions had an e

80、rror rate of less than thirty percent.Better resource planning for hospitals,especially during a crisis,leads to better patient care.This demonstrates that AI and telecom networks and data can be used to benefit society in novel and perhaps unexpected ways.Trustworthy AI-What it means for telecomCon

81、clusionJune 202316ConclusionThe benefits of AI in telecom networks are only just beginning to be leveraged but will clearly be an important and integral element in future networks.Trusting those networks requires trusting the AI,which can be achieved by following the presented guidelines.However,the

82、se present a number of challenges:Ensuring that humans retain oversight and control over AI systems,even when highly automated and operating at high speeds.Providing information about AI operations,using techniques such as XAI,while maintaining a fine balance between the privacy of models and data a

83、nd transparency Protecting the data of users and businesses,while still using it to deliver AI that benefits them Understanding how AI might affect different communities and preventing adverse impacts Making AI systems robust and safe,yet practical to train and deploy.Considering how AI affects soci

84、ety broadly,both to prevent adverse impacts and to promote beneficial uses.Trusting the AI,that is having confidence that it operates as intended and does no harm,requires diligence in addressing each of these challenges.Ericsson is working to continuously improve its AI systems,making them,and cons

85、equently their products,more trustworthy.Trustworthy AI-What it means for telecomGlossaryJune 202317GlossaryAI Artificial IntelligenceCLI Command Line InterfaceDQN Deep Q NetworkGDPR General Data Protection Regulation,https:/gdpr-info.eu/GUI Graphical User InterfaceKPI Key Performance IndicatorLIME

86、Loacl Interpretable Model-agnostic ExplanationsLMUT Linear Model U-TreeML Machine LearningMR Machine ReasoningNN Neural NetworkNOC Network Operations CenterPET Privacy Enhancing TechnologiesRET Remote Electrical TiltRL Reinforcement LearningRSRP Reference Signal Received PowerSHAP SHapley Additive e

87、xPlanationsSINR Signal to Interface plus Noise RatioSLA Service Level AgreementUE User EquipmentXAI Explainable Artificial IntelligenceXRL Explainable Reinforcement LearningTrustworthy AI-What it means for telecomReferencesJune 202318References1.https:/artificialintelligenceact.eu/2.https:/www.aepd.

88、es/sites/default/files/2019-12/ai-ethics-guidelines.pdf3.Rafia Inam,Ahmad Terra,Anusha Mujumdar,Elena Fersman,Aneta Vulgarakis,Explainable AI-How humans can trust AI.EricssonWhite Paper,April 20214.A.Terra,R.Inam,S.Baskaran,P.Batista,I.Burdick and E.Fersman,“Explainablity Methods for Identifying Roo

89、t-Cause of SLA Violation Prediction in 5G Network,”in IEEE Global Communications Conference,20205.Franco Ruggeri,Ahmad Terra,Rafia Inam,Karl-Henerik Johansson,“Evaluation of Intrinsic Explainable Reinforcement Learning in Remote Electrical Tilt Optimization”,in 8th International Congress on Informat

90、ion and Communication Technology,20236.Ahmad Terra,Rafia Inam,Elena Fersman.BEERL:Both Ends Explanations for Reinforcement Learning.Journal of Applied Sciences,Special issue“Explainable Artificial Intelligence”,Vol 12,No.21,November 20227.EU GDPR Privacy by Design,https:/gdpr-info.eu/issues/privacy-

91、by-design/8.Peter J.Rousseeuw(1987).Silhouettes:a Graphical Aid to the Interpretation and Validation of Cluster Analysis.Computational and Applied Mathematics.20:5365.doi:10.1016/0377-0427(87)90125-7.9.Dunn,J.C.(1973-09-17).A Fuzzy Relative of the ISODATA Process and Its Use in Detecting Compact Wel

92、l-Separated Clusters.Journal of Cybernetics.3(3):3257.doi:10.1080/01969727308546046.S2CID 120919314.10.A.Nikou,A.Mujumdar,V.Sundararajan,M.Orlic and A.V.Feljan,Safe RAN control:A Symbolic Reinforcement Learning Approach,2022 IEEE 17th International Conference on Control&Automation(ICCA),202211.https

93、:/ al.Development of forecast models for COVID-19 hospital admissions using anonymized and aggregated mobile network data.Sci Rep 12,17726(2022).https:/doi.org/10.1038/s41598-022-22350-613.EU General Data Protection Regulation,https:/gdpr-info.eu/tTrustworthy AI-What it means for telecomAuthorsJune

94、202319AuthorsJim Reno is a Distinguished Engineer at Ericsson,where he works on security aspects of Artificial Intelligence as applied to telecommunication systems.He has more than 40 years of industry experience in fields including system software(operating systems,networking,system management,and

95、cloud native systems),payment system security,authentication,authorization and identity management.Rafia Inam is a senior research manager at Ericsson Research and Adjunct Professor at KTH in research area Trustworthy Artificial Intelligence.She has conducted research for Ericsson for the past nine

96、years on 5G for industries,network slices and network management;AI for automation and intelligent transport systems.She specializes in automation and safety for cyber-physical systems and collaborative robots,trustworthy AI,and explainable AI.She won Ericsson Top Performance Competition 2021 on her

97、 work on AI for 5G network slice assurance,and was awarded Ericsson Key Impact Award 2020,and Key contributor award 2020.Rafia received her PhD in predictable real-time embedded software from Mlardalen University in 2014.She has co-authored 55+refereed scientific publications and 55+patent families

98、and is a program committee member,referee,and guest editor for several international conferences and journals.Trustworthy AI-What it means for telecomAuthorsJune 202320Attila Ulbert joined Ericsson in 2015 and he is currently Artificial Intelligence System Manager.In his enthusiastic journey with Ericsson,he lead the development of Ericssons AI platform,and worked on fundamental AI studies on security,trustworthiness,and industrialization.Attila has a PhD in Informatics from Etvs Lornd University.He is a marathoner.