Confidentiality is the first obstacle in a vision inspection project, and often the only one technique does not resolve. A production image shows more than a defect: it shows a part’s geometry, therefore its design, the condition of tooling, a line’s throughput, sometimes a customer name through a marking. A manufacturer hesitating to hand over those images has reasons for it.That blockage is not an administrative detail to settle after technical scoping. It determines where the work can be performed, by whom, with which tools, and it represents a cost line in its own right. A project addressing it last often discovers that part of its architecture is incompatible with what the client can accept.This article covers that constraint: what an image actually reveals, the available working architectures, what an information security framework contributes in practice, and what belongs in a contract. It extends the complete guide to defect detection.
What a defect detection image actually reveals
The first task is establishing what is genuinely sensitive in the corpus, since the answer is rarely the obvious one.The product itself
An image shows a shape, proportions, a surface finish, sometimes readable dimensions if a scale is present. On a part under development, that can be enough to reveal a technical direction. On a part in production, it is often less sensitive than assumed, since the product is visible to whoever buys it anyway.The process
This is frequently the most sensitive information and the least anticipated by technical teams. Machining marks reveal a cutting strategy, tool marks reveal equipment, surface appearance reveals a treatment. An experienced competitor infers a significant part of how a piece was produced from an image of it.The industrial context
Metadata and the image field carry information unrelated to the defect sought: timestamps allowing throughput to be reconstructed, batch numbers, machine identifiers, customer markings on the part, and sometimes people in frame.The defect rate
One point is regularly overlooked and yet decisive. A corpus of production images reveals the line’s real scrap rate, information generally treated as highly confidential and not disclosed even to customers. That disclosure does not come from individual images but from their distribution, which makes it invisible to a piece-by-piece review.Classify before protecting
Information security frameworks make classification the precondition of any protective measure, and that ordering has a practical reason.The reference standard states that information should be classified according to the organisation’s security needs in terms of confidentiality, integrity and availability, the objective being to ensure identification and understanding of protection needs. It specifies that information owners are responsible for that classification.Applied to a defect detection project, that requirement produces a useful result: not every image in a corpus sits at the same level. An image of a standard part in routine production and an image of a prototype under development do not call for the same measures, and treating them uniformly leads either to over-protecting everything, which is expensive, or under-protecting the sensitive part, which creates exposure.The practical approach is to establish with the client a three or four level classification, derive the applicable measures for each, then sort the corpus. That work takes half a day and frequently unblocks projects stuck on a blanket refusal to share.Working architectures in defect detection
Three working configurations exist, from most to least constraining, and the choice follows from the classification established.On-site annotation
Annotators work on the client’s premises, on their network, with images never leaving. That is maximum protection and it removes most objections. Its limits are capacity, restricted to what the site can host, and cost, markedly higher. It is justified on the most sensitive corpora or during scoping phases requiring close contact with quality teams.Remote access to a client environment
Images stay hosted by the client or in an environment they control, and annotators access them through a remote desktop with no local copy possible. It is the most common compromise and generally the best: it combines the capacity of a dedicated team with the absence of any transfer.Implementing it requires particular attention to what is not technically blocked. Screen capture, photographing the screen with a personal device and manual transcription of information all remain possible whatever the software lockdown. The answer lies in workstation organisation and personal commitment, not in technical configuration.Transfer to the provider’s environment
Images are transmitted and hosted by the provider. It is the most flexible configuration, the most efficient in production, and the one demanding the most guarantees. It presupposes established trust and, in practice, a verifiable demonstration of the security arrangement, which is precisely what an information security management certification provides.What a certification actually contributes
It is worth being precise here, since certifications are frequently presented as arguments without their scope being explained.What it attests
An information security management certification attests the existence of an organised and audited arrangement: a policy, a risk analysis, documented measures, assigned responsibilities, periodic controls and a management review. It attests a process, not an absolute state of security.That nuance matters and is better stated than concealed. A certified provider is not invulnerable; it has a framework making its practices verifiable, its incidents traceable and its commitments auditable. That is exactly what a manufacturer needs in order to accept a transfer, and it is markedly more than a declaration of good intent.What it changes in the relationship
The most tangible practical benefit concerns supplier assessment. The framework requires that relevant security requirements be established and agreed with each supplier according to the type of relationship, with agreements including a description of the information and access methods, its classification, each party’s obligations, subcontracting arrangements, the right of audit, periodic effectiveness reporting and end of contract procedures.In other words, the manufacturer who outsources has an obligation of their own to qualify their supplier. A certified provider supplies the material for that qualification in an already structured form, which considerably shortens the process. On projects where supplier assessment can take several months, that time saving is often the most decisive argument.What it does not replace
A certification dispenses neither with information classification, which falls to the client, nor with contract clauses specific to the project, nor with demonstrating the measures actually applied to this corpus. A provider who answers a confidentiality question by presenting a certificate without describing the concrete arrangement is dodging the question.The measures that actually matter
Beyond the framework, a few measures deliver most of the protection on an annotation project for defect detection.Access compartmentalisation
Each annotator accesses only the batches assigned to them, with a named account and logging of consultations. That granularity costs little and considerably limits the exposed surface. It has a secondary benefit noted elsewhere: it makes per-contributor performance analysis possible.No local copy
This is the most effective technical measure. A remote desktop with no drive mapping, no shared clipboard and no outbound network access removes the principal leak path. It is compatible with normal productivity provided the arrangement is correctly sized, since a slow environment pushes teams towards workarounds.The physical environment
The security framework addresses the workstation explicitly, and it is particularly relevant here. A dedicated room, no personal devices, and a clear screen policy when unattended complete the logical measures. On the most sensitive corpora, banning phones from the working area is a simple measure that closes the one vector technique cannot block.Deletion at end of engagement
The framework requires that information be deleted when no longer required, selecting an appropriate method, recording the results of deletion as evidence and, for third party services, obtaining evidence of deletion.That requirement is often handled by a vague clause. It benefits from being specified: which data is deleted, source images, annotations, backups, working environments, within what timeframe, and in what form the evidence is provided. A signed deletion report is an end of engagement deliverable on a par with the corpus.The full chain of parties
One requirement is frequently underestimated in defect detection projects: security applies to the whole chain, not only to the direct provider.The framework requires propagation of security requirements through the supply chain, validation of the conformity of delivered products and services, and obtaining assurances on security levels, with the chain covering cloud services and hosting services in particular.For an annotation project, that means an informed client will ask for the list of parties: hosting provider, annotation platform vendor, any group entities located in other countries, occasional subcontractors. Supplying that map spontaneously, rather than producing it under pressure, signals maturity.The question is all the more sensitive because market annotation platforms are frequently hosted outside the client’s country, which introduces a data localisation dimension that some sectors, notably defence and energy, do not accept.Are the images personal data
One question arises systematically in defect detection projects and rarely receives a clear answer: does personal data regulation apply to an industrial corpus.In most cases an image of a part contains no personal data and falls outside that scope. Three situations are exceptions and must be checked rather than assumed absent.The first is people appearing in frame, hands, silhouettes, reflections, badges, which happens more often than assumed at manual stations. The second is operator traceability, where an image is tied to a station and a timestamp allowing identification of who was working, which indirectly makes it data relating to a person. The third concerns logistics corpora, where labels carry names and addresses, a situation covered in another article of this cluster.The practical consequence is that a project must examine this question explicitly at scoping, and provide either for excluding the images concerned, a masking process, or a suitable legal framework. Discovering it mid-project forces the corpus to be reworked.Regulated sectors
Some contexts add defect detection constraints beyond ordinary commercial confidentiality.Defence industries impose their own classification regimes, individual clearances and nationality restrictions that rule out certain configurations outright. Aerospace and nuclear apply traceability requirements extending to providers. Some sectors mandate data localisation within national territory.These constraints cannot be worked around and are scoped upfront. A provider discovering mid-project that a nationality or localisation requirement applies is in an impasse. The question to ask in the first exchange is simple: does the corpus fall under a particular regulated regime, and which.What unblocks a reluctant client
Reluctance to share production images for defect detection is rational, and arguing frontally works badly. Four approaches give better results.Classification by level, discussed above, turns a blanket refusal into a differentiated decision. A client refusing to share their corpus often accepts sharing its least sensitive part, which is enough to start.Starting on site, even for a few weeks, builds trust through direct observation of practices. It costs more but unblocks situations no documentation unblocks.Cropping images, where the task permits, considerably reduces the information transmitted. A crop centred on an inspection zone reveals neither the part’s full geometry nor its context.Finally, transparency about the arrangement’s limits is more effective than overselling it. A provider who explains what their configuration blocks and what it does not, and who proposes organisational measures for the rest, earns more credit than a claim of total security nobody believes.The cost of security in a defect detection project
These measures carry a price that should appear in the costing rather than being absorbed silently.Three cost lines stand out. Infrastructure, compliant hosted environment, remote desktops, compartmentalisation, logging, which is a fixed charge independent of volume. Organisation, dedicated room, procedures, training, internal audits, which is also largely fixed. And productivity loss from the constraints, since remote access on large images is slower than local work, which translates into a real unit cost premium.That structure has a consequence: security weighs proportionally more on small projects. A corpus of a few thousand images in a constrained environment can see its unit cost double relative to the same work in a standard configuration, which should be stated at quotation rather than discovered.Security as a selection criterion
From the point of view of a manufacturer selecting a defect detection provider, knowing what to ask is more useful than a list of certifications. Six questions distinguish a real arrangement from a narrative.Where will my images be physically hosted, and in which country. Who exactly will have access, and under what compartmentalisation. Who are all the parties in the chain, hosting provider and platform vendor included. What happens at end of engagement, and in what form is deletion evidence provided. What measures cover the vectors technique cannot block, notably screen capture. And can I audit, under what concrete arrangements.A provider able to answer all six precisely has a tested arrangement. A provider who answers all six by producing a certificate has probably not thought about the project, and that difference shows within minutes of conversation. It is also, from the provider’s side, the best preparation for a supplier assessment: answering them spontaneously in a proposal considerably shortens the evaluation.The contract
Five points must appear in a defect detection contract, and their absence is a scoping failure rather than a legal oversight.The purpose and scope of use, meaning what the provider is authorised to do with the images and what they are not. The retention period and deletion procedure, with the nature of the evidence supplied. The list of parties and the conditions for adding a new one. The right of audit, with its practical arrangements, otherwise the clause stays theoretical. And the incident notification procedure, with its timeframe.A sixth point deserves explicit and separate negotiation: corpus reuse. Three regimes are possible and must be distinguished. Use strictly limited to the project with full deletion. Retention by the client alone, the provider keeping nothing. Or authorisation of derived use by the provider, for instance to improve their own generic models. That third regime is legitimate but must be requested and granted, never assumed.Security and annotation quality
One tension in defect detection projects deserves naming because it is real and usually left unspoken: constrained environments degrade annotation quality if they are badly implemented.Three mechanisms operate. Latency, since an annotator waiting several seconds for each image works faster and less carefully to compensate. Restricted tooling, since a locked-down environment sometimes prevents installing the tool best suited to the task, forcing a compromise. And isolation, since annotators separated from the rest of the team by a physical or network barrier receive feedback and protocol updates more slowly.None of these is inevitable, and all three are cheaper to prevent than to correct. Sizing the remote environment for the real image volume rather than for a demonstration, validating the tooling before committing to a configuration, and organising deliberate feedback channels into the secured zone are the three countermeasures. They cost something and should appear in the costing, since the alternative is a corpus that is secure and mediocre, which serves nobody.The most common mistakes
These failures recur often enough across industrial projects that naming them is usually enough to avoid them.- Addressing confidentiality after technical scoping rather than before.
- Applying a uniform protection level to a heterogeneous corpus.
- Overlooking what the corpus distribution reveals, notably the real scrap rate.
- Relying on software lockdown to block photographing the screen.
- Presenting a certification without describing the concrete arrangement applied.
- Omitting the full map of parties, hosting provider and platform included.
- Providing a deletion clause without specifying scope or evidence.
- Assuming authorisation for derived use of the corpus rather than negotiating it.
- Discovering a nationality or localisation constraint mid-project.
- Absorbing security costs into the unit price without showing them.
